Communication method and system based on fusion autoencoder and generative adversarial network

Abstract

The present application discloses a communication method and system based on a fusion autoencoder and a generative adversarial network, the method comprising: encoding an input signal through an encoding network of an autoencoder to obtain a true coded signal and modulate it, and quantizing the received signal after passing through a channel to obtain a quantized signal; thereby fitting a random signal to a false coded signal that approximates a true coded signal through a generator of a generative adversarial network; fusing the decoding network of the autoencoder with the discriminator of the generative adversarial network into an initial network, and training the initial network based on a preset weighted loss function and a training strategy with a turning point, obtaining a fusion network that judges the true or false of a coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal. Thus, the problems in the related art that accurate channel information cannot be obtained in a communication system with a complex channel, the complexity of system implementation is increased, and the reliability and accuracy of the communication system are reduced.

Description

Technical Field

[0001] The present application relates to the field of intelligent communication technology, and in particular to a communication method and system based on a fusion autoencoder and a generative adversarial network. Background Art

[0002] In the related art, a communication system can obtain channel information through an accurate channel model and classify the input / output signals according to the mathematical characteristics of the input / output signals and the mathematical characteristics of the relationship between the input / output signals to improve the communication quality.

[0003] However, in related technologies, since communication is more dependent on channel models, accurate channel information cannot be obtained in communication systems with more complex channels, which increases the complexity of system implementation and reduces the reliability and accuracy of the communication system, which urgently needs to be solved. Summary of the invention

[0004] This application is based on the following problems and understandings made by the inventor:

[0005] With the maturity of AI (Artificial Intelligence) technology, intelligent communication applying AI to wireless communication has become one of the hot research directions. It has achieved excellent performance in wireless communication fields such as channel estimation, channel feedback, beamforming, and signal detection. It is considered to be one of the trend technologies of 6G. DL (Deep Learning), as an important subset of AI, has received widespread attention from academia and industry. Different from traditional communications that rely more on accurate channel models to improve communication quality, DL's powerful adaptive nonlinear fitting ability can learn the complex relationships between variables, especially those variables that are difficult to accurately describe with mathematical models.

[0006] However, for communication systems with more complex channels, such as time-selective channels, underwater optical channels, free-space optical channels, power line channels, etc., it is difficult for traditional communication schemes to obtain accurate channel information, thereby reducing the reliability of traditional communications. At the same time, in order to reduce power consumption, the quantization introduced at the receiving end will also bring nonlinearity to the communication system, further affecting the accuracy of signal transmission. Deep learning can fit the quantized signal that is interfered by factors such as multipath fading into the transmitted signal based on its nonlinear fitting ability, without the need for channel information, and achieve better performance with lower computational complexity. Therefore, for complex communication environments, it is very important to develop communication methods based on deep learning.

[0007] The present application provides a communication method and system based on a fusion of an autoencoder and a generative adversarial network to solve the problems in the related art that, due to the communication's greater reliance on a channel model, accurate channel information cannot be obtained in a communication system with a more complex channel, which increases the complexity of the system implementation and reduces the reliability and accuracy of the communication system.

[0008] The first aspect of the present application provides a communication method based on a fusion of an autoencoder and a generative adversarial network, comprising the following steps: encoding an input signal through an encoding network of an autoencoder to obtain a true coded signal; modulating the true coded signal, and quantizing the received signal after passing through a channel to obtain a quantized signal; based on the quantized signal, fitting a random signal into a false coded signal that approximates the true coded signal through a generator of a generative adversarial network; fusing the decoding network of the autoencoder and the discriminator of the generative adversarial network into an initial network, and training the initial network based on a preset weighted loss function and a training strategy with a turning point, to obtain a fusion network that judges the truth or falsity of a coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal.

[0009] Optionally, in one embodiment of the present application, the input signal is encoded by the encoding network of the automatic encoder to obtain a true encoded signal, including: generating a corresponding symbol according to each preset bit information of the bit sequence, and mapping the symbol into corresponding encoding information; performing normalization processing based on the encoding information to obtain the true encoded signal.

[0010] Optionally, in one embodiment of the present application, the modulating the true coded signal includes: modulating the true coded signal based on OOK (On-Off Keying, binary amplitude keying), BPSK (Binary Phase Shift Keying, binary phase shift keying), QAM (Quadrature amplitude modulation, orthogonal amplitude modulation) or OFDM (Orthogonal Frequency Division Multiplexing, orthogonal frequency division multiplexing) modulation; and sending the modulated true coded signal using a transmitting end.

[0011] Optionally, in one embodiment of the present application, the quantizing of the received signal after passing through the channel to obtain the quantized signal includes: based on an adaptive quantization method, using a quantization function quantization threshold as a trainable parameter to adaptively adjust to an optimal threshold.

[0012] Optionally, in one embodiment of the present application, the quantization function is an approximate unit step function.

[0013] Optionally, in one embodiment of the present application, the training strategy with a turning point includes: before the turning point, the loss function weight of the autoencoder is greater than a preset threshold, after the turning point, the parameters of the encoding network are fixed, and the loss function weights of the autoencoder and the generative adversarial network are the same.

[0014] The second aspect of the present application provides a communication system based on a fusion of an autoencoder and a generative adversarial network, including: an encoding module, used to encode an input signal through the encoding network of the autoencoder to obtain a true coded signal; a first processing module, used to modulate the true coded signal, and quantize the received signal after passing through the channel to obtain a quantized signal; a fitting module, used to fit a random signal into a false coded signal that approximates the true coded signal through a generator of a generative adversarial network based on the quantized signal; a second processing module, used to fuse the decoding network of the autoencoder and the discriminator of the generative adversarial network into an initial network, and train the initial network based on a preset weighted loss function and a training strategy with a turning point, to obtain a fused network that judges the truth or falsity of the coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal.

[0015] Optionally, in one embodiment of the present application, the encoding module includes: a generation unit, used to generate corresponding symbols according to each preset bit information of the bit sequence, and map the symbols into corresponding encoding information; a processing unit, used to perform normalization processing based on the encoding information to obtain the true encoding signal.

[0016] Optionally, in one embodiment of the present application, the first processing module includes: a modulation unit, used to modulate the true coded signal based on binary amplitude keying OOK, two-phase phase shift keying BPSK, orthogonal amplitude modulation QAM or orthogonal frequency division multiplexing modulation OFDM modulation; a sending unit, used to send the modulated true coded signal using a transmitting end.

[0017] Optionally, in one embodiment of the present application, the first processing module is further used to use a quantization function to quantize a threshold value as a trainable parameter based on an adaptive quantization method, so as to adaptively adjust to an optimal threshold value.

[0018] Optionally, in one embodiment of the present application, the quantization function is an approximate unit step function.

[0019] Optionally, in one embodiment of the present application, the training strategy with a turning point includes that before the turning point, the loss function weight of the autoencoder is greater than a preset threshold, and after the turning point, the parameters of the encoding network are fixed, and the loss function weights of the autoencoder and the generative adversarial network are the same.

[0020] The third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the communication method based on a fusion autoencoder and a generative adversarial network as described in the above embodiment.

[0021] The fourth aspect embodiment of the present application provides a computer-readable storage medium, which stores a computer program, which, when executed by a processor, implements the above-mentioned communication method based on a fusion autoencoder and a generative adversarial network.

[0022] The embodiment of the present application can encode the input signal through the encoding network of the autoencoder to obtain a true coded signal and modulate it, and quantize the received signal after passing through the channel to obtain a quantized signal, so that the random signal is fitted into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network, and the decoding network of the autoencoder and the discriminator of the generative adversarial network are fused into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, and a fusion network that judges the true or false of the coded signal to play a game with the generator and decodes any coded signal to restore it to a transmission signal is obtained, so that accurate channel information can be obtained in a communication system with a complex channel, and the reliability and accuracy of the communication system are effectively improved. Thus, the problem that the communication in the related art is more dependent on the channel model, resulting in the inability to obtain accurate channel information in a communication system with a complex channel, increasing the complexity of system implementation, and reducing the reliability and accuracy of the communication system is solved.

[0023] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0025] Figure 1 A flow chart of a communication method based on a fusion autoencoder and a generative adversarial network provided according to an embodiment of the present application;

[0026] Figure 2 A schematic diagram of an approximate unit step function according to a specific embodiment of the present application;

[0027] Figure 3 A schematic diagram of the system structure built for the training process of a specific embodiment of the present application;

[0028] Figure 4A schematic diagram of a system structure built for a test process of a specific embodiment of the present application;

[0029] Figure 5 It is a structural schematic diagram of a communication system based on a fusion autoencoder and a generative adversarial network according to an embodiment of the present application;

[0030] Figure 6 Schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0031] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0032] The following is a description of a communication method and system based on a fusion autoencoder and a generative adversarial network according to an embodiment of the present application with reference to the accompanying drawings. In view of the problem that in the related art mentioned in the background technology center, since communication is more dependent on the channel model, accurate channel information cannot be obtained in a communication system with a complex channel, the complexity of the system implementation is increased, and the reliability and accuracy of the communication system are reduced, the present application provides a communication method based on a fusion autoencoder and a generative adversarial network, in which the input signal can be encoded by the encoding network of the autoencoder to obtain a true coded signal and modulate it, and the received signal after passing through the channel is quantized to obtain a quantized signal, so that the random signal is fitted to a false coded signal approximating the true coded signal by the generator of the generative adversarial network, and the decoding network of the autoencoder is fused with the discriminator of the generative adversarial network into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, so as to obtain a fusion network that judges the true or false of the coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal, so that accurate channel information can be obtained in a communication system with a complex channel, effectively improving the reliability and accuracy of the communication system. This solves the problem in the related art that, due to the communication being more dependent on the channel model, accurate channel information cannot be obtained in a communication system with a more complex channel, which increases the complexity of system implementation and reduces the reliability and accuracy of the communication system.

[0033] Specifically, Figure 1 A flow chart of a communication method based on a fusion autoencoder and a generative adversarial network provided in an embodiment of the present application.

[0034] like Figure 1 As shown, the communication method based on the fusion of the automatic encoder and the generation of the adversarial network includes the following steps:

[0035] In step S101, the input signal is encoded by the encoding network of the autoencoder to obtain a true encoded signal.

[0036] It can be understood that the embodiment of the present application can encode the input signal through the encoding network of the automatic encoder in the following steps and obtain a true encoded signal, wherein the encoding network can be composed of multiple layers of fully connected layers, and the dimension of the obtained true encoded signal is higher than the number of bits carried by the signal, thereby increasing redundant bits to improve the reliability of communication.

[0037] Optionally, in one embodiment of the present application, the input signal is encoded by the encoding network of the automatic encoder to obtain a true encoded signal, including: generating corresponding symbols according to each preset bit information of the bit sequence, and mapping the symbols into corresponding encoding information; performing normalization processing based on the encoding information to obtain a true encoded signal.

[0038] During the actual implementation process, the embodiment of the present application can treat each bit of information in the bit sequence as a group, and generate a corresponding group symbol, and map the symbol to the corresponding coding information, perform normalization processing based on the coding information, and obtain a true coded signal, thereby improving the reliability of transmission by increasing redundancy, and effectively improving the anti-interference ability of uncoded bit stream information.

[0039] It should be noted that the preset bit information is set by those skilled in the art according to actual conditions and is not specifically limited here.

[0040] In step S102, the true coded signal is modulated, and the received signal after passing through the channel is quantized to obtain a quantized signal.

[0041] It can be understood that the embodiments of the present application can modulate the true coded signal and quantize the received signal after passing through the channel in the following steps to obtain a quantized signal. That is to say, the true coded signal is modulated and sent, and reaches the receiving end through a non-ideal channel, and is quantized at the receiving end to obtain a quantized signal, thereby improving the feasibility of the communication system while improving the reliability of the non-ideal channel and quantization on the communication.

[0042] Optionally, in one embodiment of the present application, the true coded signal is modulated, including: modulating the true coded signal using OOK, BPSK, QAM or OFDM modulation; and sending the modulated true coded signal using a transmitting end.

[0043] As a possible implementation method, the embodiments of the present application can modulate the true coded signal. For example, the modulation method may include but is not limited to OOK, BPSK, QAM or OFDM, and use a non-ideal channel to send the modulated true coded signal. For example, the non-ideal channel includes but is not limited to a wireless Rayleigh channel, a wireless Ricean channel, a time selective channel, an underwater optical channel, a free space optical channel, a power line channel, etc., thereby achieving accurate channel information in a communication system with more complex channels, reducing the complexity of system implementation.

[0044] Optionally, in one embodiment of the present application, the received signal after passing through the channel is quantized to obtain a quantized signal, including: based on an adaptive quantization method, using a quantization function quantization threshold as a trainable parameter to adaptively adjust to an optimal threshold.

[0045] In some embodiments, the quantization in the embodiments of the present application can be based on an adaptive quantization method during the training process, using the quantization function quantization threshold as a trainable parameter, and adaptively adjusted to the optimal threshold. In the test process after the training is completed, a conventional quantization method can be used, so that the quantization threshold is fixed to the optimal threshold learned during the test process, and preferably, for example, it can be a one-bit quantization, wherein the quantization function of the adaptive quantization method in the training process is an approximate unit step function, and the quantization function of the conventional quantization method in the test process is a unit step function.

[0046] Optionally, in one embodiment of the present application, the quantization function is an approximate unit step function.

[0047] Among them, the quantization function is an approximate unit step function, and the approximate unit step function is:

[0048]

[0049] Among them, f sigmoid is the approximate unit step function and is the scaling parameter.

[0050] In step S103, based on the quantized signal, the random signal is fitted into a false coded signal that approximates the true coded signal through a generator of a generative adversarial network.

[0051] It can be understood that the embodiments of the present application can be based on the quantized signal, use the quantized signal as conditional information, and fit the random signal into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network, wherein the generator is equivalent to an adaptive channel equalizer, which can be composed of multiple layers of one-dimensional convolutional layers and fully connected layers. The random noise can be fitted to make it close to the transmitted signal, thereby realizing channel self-equalization when the channel state information is unknown, thereby reducing the complexity of the communication system implementation.

[0052] In step S104, the decoding network of the autoencoder and the discriminator of the generative adversarial network are fused into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, so as to obtain a fused network that can judge the truth or falsity of a coded signal to compete with the generator and decode any coded signal to restore it to a transmission signal.

[0053] It can be understood that the embodiment of the present application can fuse the decoding network of the autoencoder and the discriminator of the generative adversarial network into one network, namely the initial network, and train the initial network based on the weighted loss function and the training strategy with turning points. The fused network can judge the truth or falsity of the encoded signal to compete with the generator, and can decode any encoded signal to restore it to the fused network of the transmission signal, thereby realizing channel self-balancing when the channel state information is unknown, effectively improving the reliability of communication due to non-ideal channels and quantization, reducing the complexity of system implementation, and improving the system communication performance.

[0054] Among them, the fusion network can simultaneously realize the discriminator to judge the truth or falseness of the coded signal, and the decoding network to decode the coded signal. It can be composed of multiple layers of one-dimensional convolutional layers and fully connected layers. The input is a true coded signal or a false coded signal, and the quantized signal is used as conditional information. The output is the decoded signal and the probability of judging whether the coded signal is true.

[0055] Among them, the weighted loss function is the weighted sum of the loss function of the conventional generative adversarial network and the loss function of the autoencoder, and the weighted values ​​of the two can be adjusted during the training process.

[0056] Among them, in one embodiment of the present application, the training strategy with a turning point includes: before the turning point, the loss function weight of the autoencoder is greater than a preset threshold, after the turning point, the parameters of the encoding network are fixed, and the loss function weights of the autoencoder and the generative adversarial network are the same.

[0057] In the actual implementation process, the training strategy with turning points has a higher weight of the loss function of the autoencoder before the turning point, and focuses more on the training of the communication reliability of the transceiver. After the turning point, the parameters of the encoding network are fixed, which is more conducive to the training of the generative adversarial network. At the same time, the loss function weights of the autoencoder and the generative adversarial network are the same.

[0058] It should be noted that the preset threshold is set by those skilled in the art according to actual conditions and is not specifically limited here.

[0059] The working principle of the embodiments of the present application is described in detail below with reference to some specific embodiments. Specific embodiment 1

[0061] For example, the embodiment of the present application may use OOK modulation to modulate the true coded signal, wherein the channel is an underwater optical channel, and one-bit quantization is used to reduce energy consumption. The specific steps are as follows:

[0062] Step S11, the embodiment of the present application encodes the input signal through the encoding network of the automatic encoder, and the obtained encoded signal is called a true encoded signal.

[0063] Specifically, the embodiment of the present application can treat each k bits of information in the bit sequence as a group to form a symbol s i , then the number of symbol types is M = 2 k , the input of the encoding network is s i The one-hot representation is that the i-th value of the M-dimensional zero vector is set to 1, where the encoding network consists of several fully connected layers to convert the symbol s i Mapping to coded signal And normalize it at the end. Taking the j-th fully connected layer as an example, the process can be expressed as:

[0064] h j =f ReLU (W j h j-1 +b j ),

[0065] Among them, h j , W j 、b j are the feature vector, weight matrix and offset of the jth layer, respectively, ReLU An activation function for nonlinear fitting.

[0066] In addition, since the uncoded bit stream information has poor anti-interference ability, the coding network can improve the reliability of transmission by increasing redundancy.

[0067] In step S12, the embodiment of the present application can send the true coded signal after OOK modulation, and reach the receiving end through the underwater optical channel. The receiving end performs one-bit quantization and obtains a quantized signal, and uses the quantized signal as conditional information to fit the random signal into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network.

[0068] It should be noted that OOK modulation is a binary modulation method, which is widely used in optical communications because of its simplicity and ease of implementation. Underwater wireless optical communication is mainly affected by absorption and scattering. The collision of photons with other particles will cause the loss of photon energy, and the deviation of the propagation direction of photons will cause the number of photons detected by the receiver to decrease, thereby causing power attenuation. At the same time, due to multiple scattering, photons may arrive at the receiver in different time slots, resulting in multipath effects. These interferences are difficult to accurately describe with mathematical formulas. Therefore, the fusion of autoencoders and generative adversarial networks has significant advantages.

[0069] Furthermore, one-bit quantization at the receiving end can reduce power consumption. The quantization value can be determined by comparing the received signal y with the threshold t. Due to the influence of multipath fading in the underwater optical channel, the optimal threshold cannot be determined. Therefore, it can be used as a trainable parameter and adaptively adjusted to the optimal threshold that can minimize the system bit error rate during the training process. The threshold t is optimized, and the quantization function is smooth and differentiable, so that the gradient of the loss function can be back-propagated. Therefore, the conventional unit step function is not applicable, and the embodiment of the present application adopts the following approximate unit step function, namely:

[0070]

[0071] Among them, γ is the scaling parameter.

[0072] In addition, when the scaling parameter γ is a large non-negative real number, the effect of the function is very close to the unit step function and is smooth and differentiable.

[0073] like Figure 2 As shown, Figure 2 This is an example of an approximate unit step function in an embodiment of the present application, wherein the threshold t is set to 2, and γ is 10, 20, and 100 from small to large. It can be obtained that the larger the γ, the closer the function is to the unit step function. Therefore, during training, the approximate unit step function can help the system achieve the goal of adaptively optimizing the threshold t. During testing, in order to simulate actual applications, a conventional unit step function can be used as the quantization function, and the threshold t is fixed to the optimal threshold obtained during the training process.

[0074] Specifically, the quantized signal q obtained after one-bit quantization in the embodiment of the present application includes the true coded signal e r Therefore, it can be used as conditional information and input into the generator together with the random signal to obtain the approximate true coding signal e r The pseudo-coded signal e f ,To achieve the goal of channel equalization, the generator contains multiple layers of one-dimensional convolutional layers to learn the impact of underwater optical channels on the transmitted signals. Then, in order to simulate the coded signal e r, we can add several fully connected layers after the one-dimensional convolutional layer and normalize it at the end, so that the pseudo-coded signal e obtained after generator equalization is f It will be input into the decoding network for decoding, which is more conducive to the symbol s i reconstruction.

[0075] Step S13, the embodiment of the present application can merge the decoding network of the autoencoder and the discriminator of the generative adversarial network into one network, which can not only judge the authenticity of the encoded signal to compete with the generator, but also decode the encoded signal and restore it to the input signal.

[0076] Among them, the discriminator can convert the true encoded signal e generated by the encoding network r The pseudo-coded signal e generated by the generator f The decoding network aims to recover the symbol s from the encoded signal. i , it can be seen that the discriminator and the decoding network can input the same information, and at the same time lay the foundation for the discriminator and the decoding network to share the same network.

[0077] Furthermore, from the output of the discriminator and the decoding network, the discriminator solves a binary classification problem, while the decoding network solves an M-classification problem, where each category represents a symbol s i In order to enable the discriminator and the decoding network to share the same network DD to reduce the overall size of the system and avoid mutual interference caused by different output targets of the two, the output dimension of DD becomes M+1, and the first M elements are responsible for M classification tasks to recover the symbol s i , the last element is used to judge the true or false of the coded signal. When the coded signal is true, the coded signal e r When , the value of the last element is expected to be close to 1. When the coded signal is a false coded signal e f , then we expect the last element to have a value close to 0.

[0078] In addition, in order to enable DD to distinguish true and false coded signals and determine whether the coded signal corresponds to the correct quantized signal, the quantized signal q is also used as part of the shared network input. At the same time, the quantized signal q also contains the symbol s i information, which is helpful for symbol s i reconstruction.

[0079] Step S14, the embodiment of the present application is based on a weighted loss function and a training strategy with turning points, which ensures that the end-to-end training of the autoencoder and the game training of the generative adversarial network can be carried out simultaneously, and the performance of the trained system is tested.

[0080] In summary, DD can not only determine the authenticity of the coded signal, but also reconstruct it into the transmitted symbol. Therefore, the loss function is the weighted sum of the two objectives, namely:

[0081]

[0082] Among them, α is the weight coefficient that weighs the discrimination ability and symbol reconstruction accuracy, The goal of distinguishing true from false is as shown below.

[0083] Distinguishing true from false goals The expression is:

[0084]

[0085] Among them, DD b (·) gives the probability that DD judges the coded signal as a true coded signal.

[0086] and Recover the transmitted symbol s from the true and false coded signals respectively i The goal is expressed as:

[0087]

[0088]

[0089] Among them, DD m (·) is the reconstruction of the transmitted symbol.

[0090] For the generator, it is possible to generate sufficiently realistic encoded signal confusion DD. The goal can be expressed as:

[0091]

[0092] Furthermore, the generator can determine which type of symbol s the pseudo-encoded signal should be reconstructed into by DD i , so it can also be obtained by Description, the loss function of the generator is also the weighted sum of the above two parts, which can be expressed as:

[0093]

[0094] Among them, β is the weight coefficient.

[0095] like Figure 3 As shown, during training, the symbol s sent i First, the true coded signal e is obtained through the coding network r Then, the pseudo-coded signal e is obtained through binarization, OOK modulation, underwater optical channel, one-bit quantization and generator. f, DD can distinguish true and false coded signals and restore them to symbols s according to the quantized signal q i During the training process, the parameters of the encoding network and the discriminator parameters are optimized as a group, so that the end-to-end joint optimization of the autoencoder can be achieved. At the same time, the gradient of the encoding network parameters can be obtained only through DD, so that the binarization process of OOK modulation will not affect the back propagation of the encoding network gradient. The generator parameters are another group, and the two groups of parameters are trained alternately and cyclically, thereby ensuring the game optimization of the generative adversarial network.

[0096] In addition, the training process contains a key turning point, before which more emphasis is placed on training the accuracy of symbol recovery, i.e. and The weight is higher, It can make the communication system obtain an optimal standard based on the ideal channel. It can prevent overfitting. After the turning point, the parameters of the coding network will be fixed, so that the generator and DD have fixed conditional information, that is, the quantization signal q, which is more conducive to the stability of training. After the turning point, the weights become 0.5, so that the fitting ability of the generator is gradually improved, and the gradual approach of the true and false coding signals will further improve the accuracy of DD in recovering the sent symbols.

[0097] like Figure 4 As shown in the figure, the test process has a simpler structure than the training process because there is no need for DD to distinguish true and false coding signals. The true coding signal from the encoder is e r It will not directly enter DD, but will only go through OOK modulation, underwater optical channel and one-bit quantization in sequence to obtain the quantized signal q. The trained generator obtains the approximate true coded signal e based on the quantized signal q. r The pseudo-coded signal e f , and enter DD for symbol s i recovery. Specific embodiment 2

[0099] For another example, the embodiment of the present application may use BPSK modulation to modulate the true coded signal, wherein the channel is a wireless Rayleigh channel, and one-bit quantization is used to reduce energy consumption. The method includes the following steps:

[0100] Step S21, encode the input signal through the encoding network of the automatic encoder, and the obtained encoded signal is called the true encoded signal.

[0101] Specifically, the embodiment of the present application can treat each k bits of information in the bit sequence as a group to form a symbol s i , then the number of symbol types is M = 2 k , the input of the encoding network is s iThe one-hot representation is that the i-th value of the M-dimensional zero vector is set to 1, where the encoding network consists of several fully connected layers to convert the symbol s i Mapping to coded signal And normalize it at the end. Taking the j-th fully connected layer as an example, the process can be expressed as:

[0102] h j =f ReLU (W j h j-1 +b j ),

[0103] Among them, h j , W j 、b j are the feature vector, weight matrix and offset of the jth layer, respectively, ReLU An activation function for nonlinear fitting.

[0104] In addition, since the uncoded bit stream information has poor anti-interference ability, the coding network can improve the reliability of transmission by increasing redundancy.

[0105] In step S22, the embodiment of the present application can perform BPSK modulation on the true coded signal and then send it to the receiving end through the wireless Rayleigh channel. The receiving end performs one-bit quantization to obtain a quantized signal, and uses the quantized signal as conditional information to fit the random signal into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network.

[0106] Furthermore, one-bit quantization at the receiving end can reduce power consumption. The quantization value is determined by comparing the received signal y with the threshold t. Due to channel interference, the optimal threshold cannot be determined. Therefore, it is used as a trainable parameter. During the training process, it is adaptively adjusted to the optimal threshold that can minimize the system bit error rate, and the threshold t is optimized. The quantization function is smooth and differentiable, so that the gradient of the loss function can be back-propagated. Therefore, the conventional unit step function is not applicable, and the embodiment of the present application adopts the following approximate unit step function, namely:

[0107]

[0108] Among them, γ is the scaling parameter.

[0109] In addition, when the scaling parameter γ is a large non-negative real number, the effect of the function is very close to the unit step function and is smooth and differentiable. Therefore, during training, the approximate unit step function can help the system achieve the goal of adaptively optimizing the threshold t. During testing, in order to simulate actual applications, a conventional unit step function is used as the quantization function, and the threshold t is fixed to the optimal threshold obtained during the training process.

[0110] Specifically, the quantized signal q obtained after one-bit quantization in the embodiment of the present application includes the true coded signal e r Therefore, it can be used as conditional information and input into the generator together with the random signal to obtain the approximate true coding signal e r The pseudo-coded signal e f , to achieve the goal of channel equalization, the generator contains multiple layers of one-dimensional convolutional layers to learn the impact of the channel on the transmitted signal. Then, in order to simulate the coded signal e r , we can add several fully connected layers after the one-dimensional convolutional layer and normalize it at the end, so that the pseudo-coded signal e obtained after generator equalization is f It will be input into the decoding network for decoding, which is more conducive to the symbol s i reconstruction.

[0111] In step S23, the embodiment of the present application integrates the decoding network of the autoencoder and the discriminator of the generative adversarial network into one network, which can not only judge the authenticity of the encoded signal to compete with the generator, but also decode the encoded signal and restore it to the input signal.

[0112] Among them, the role of the discriminator is to convert the true encoding signal e generated by the encoding network r The pseudo-coded signal e generated by the generator f The decoding network aims to recover the symbol s from the encoded signal. i , it can be seen that the discriminator and the decoding network can input the same information, and at the same time lay the foundation for the discriminator and the decoding network to share the same network.

[0113] Furthermore, from the output of the discriminator and the decoding network, the discriminator solves a binary classification problem, while the decoding network solves an M-classification problem, where each category represents a symbol s i In order to enable the discriminator and the decoding network to share the same network DD to reduce the overall size of the system and avoid mutual interference caused by different output targets of the two, the output dimension of DD becomes M+1, and the first M elements are responsible for M classification tasks to recover the symbol s i , the last element is used to judge the true or false of the coded signal. When the coded signal is true, the coded signal e r When , the value of the last element is expected to be close to 1. When the coded signal is a false coded signal e f , then we expect the last element to have a value close to 0.

[0114] In addition, in order to enable DD to not only distinguish true and false coded signals, but also to determine whether the coded signal corresponds to the correct quantized signal, the quantized signal q is also used as part of the shared network input. At the same time, the quantized signal q also contains the symbol s iinformation, which is helpful for symbol s i reconstruction.

[0115] Step S24, the embodiment of the present application is based on a weighted loss function and a training strategy with a turning point, which ensures that the end-to-end training of the autoencoder and the game training of the generative adversarial network can be carried out simultaneously, and the performance of the trained system is tested.

[0116] In summary, DD can not only determine the authenticity of the coded signal, but also reconstruct it into the transmitted symbol. Therefore, the loss function is the weighted sum of the two objectives, namely:

[0117]

[0118] Among them, α is the weight coefficient that weighs the discrimination ability and symbol reconstruction accuracy, The goal of distinguishing true from false is as shown below.

[0119] Distinguishing true from false goals The expression is:

[0120]

[0121] Among them, DD b (·) gives the probability that DD judges the coded signal as a true coded signal.

[0122] and Recover the transmitted symbol s from the true and false coded signals respectively i The goal is expressed as:

[0123]

[0124]

[0125] Among them, DD m (·) is the reconstruction of the transmitted symbol.

[0126] For the generator, it is possible to generate sufficiently realistic encoded signals to confuse DD. The goal is expressed as:

[0127]

[0128] Furthermore, the generator can determine which type of symbol s the pseudo-encoded signal should be reconstructed into by DD i , so it can also be obtained by Description, the loss function of the generator is also the weighted sum of the above two parts, which can be expressed as:

[0129]

[0130] Among them, β is the weight coefficient.

[0131] During training, the symbol s sent i First, the true coded signal e is obtained through the coding network r Then, the pseudo-coded signal e is obtained by BPSK modulation, wireless Rayleigh channel, one-bit quantization and generator. f , DD is used to distinguish true and false coded signals and restore them to symbols s according to the quantized signal q i During the training process, the parameters of the encoding network and the discriminator parameters are optimized as a group, so that the end-to-end joint optimization of the autoencoder can be achieved. At the same time, the gradient of the encoding network parameters can be obtained only through DD, so that the binarization process of BPSK modulation will not affect the back propagation of the encoding network gradient. The generator parameters are another group, and the two groups of parameters are trained alternately and cyclically, thereby ensuring the game optimization of the generative adversarial network.

[0132] In addition, the training process contains a key turning point, before which more emphasis is placed on training the accuracy of symbol recovery, i.e. and The weight is higher, It can make the communication system obtain an optimal standard based on the ideal channel. It can prevent overfitting. After the turning point, the parameters of the coding network will be fixed, so that the generator and DD have fixed conditional information, that is, the quantization signal q, which is more conducive to the stability of training. After the turning point, the weights become 0.5, so that the fitting ability of the generator is gradually improved, and the gradual approach of the true and false coding signals will further improve the accuracy of DD in recovering the sent symbols.

[0133] The structure of the test process is simpler than that of the training process, because there is no need for DD to distinguish true and false encoded signals. The true encoded signal from the encoder is r It will not directly enter DD, but will only go through BPSK modulation, wireless Rayleigh channel and one-bit quantization in sequence to obtain the quantized signal q. The trained generator obtains the approximate true coded signal e based on the quantized signal q. r The pseudo-coded signal e f , and enter DD for symbol s i recovery. Specific embodiment three

[0135] For another example, the embodiment of the present application can use 16QAM modulation to modulate the true coded signal, wherein the channel is a wireless Rice channel, and two-bit quantization is used to reduce energy consumption. The method includes the following steps:

[0136] Step 31, the embodiment of the present application encodes the input signal through the encoding network of the automatic encoder, and the obtained encoded signal is called a true encoded signal.

[0137] Specifically, the embodiment of the present application can treat each k bits of information in the bit sequence as a group to form a symbol s i , then the number of symbol types is M = 2 k , the input of the encoding network is s i The one-hot representation is that the i-th value of the M-dimensional zero vector is set to 1, where the encoding network consists of several fully connected layers to convert the symbol s i Mapping to coded signal And normalize it at the end. Taking the j-th fully connected layer as an example, the process can be expressed as:

[0138] h j =f ReLU (W j h j-1 +b j ),

[0139] Among them, h j , W j 、b j are the feature vector, weight matrix and offset of the jth layer, respectively, ReLU An activation function for nonlinear fitting.

[0140] In addition, since the uncoded bit stream information has poor anti-interference ability, the coding network can improve the reliability of transmission by increasing redundancy.

[0141] Step S32, the embodiment of the present application can send the true coded signal after 16QAM modulation, and reach the receiving end through an infinite Ricean channel. The receiving end performs two-bit quantization to obtain a quantized signal, and uses the quantized signal as conditional information to fit the random signal into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network.

[0142] Furthermore, two-bit quantization at the receiving end can reduce power consumption. The quantization value is obtained by dividing the received signal y by the threshold t i (i=1,2) to determine the optimal threshold. However, due to channel interference, it is impossible to determine the optimal threshold. Therefore, it can be used as a trainable parameter. During the training process, it is adaptively adjusted to the optimal threshold that can minimize the system bit error rate. The threshold t can be optimized. i , and the quantization function is smooth and differentiable, so that the gradient of the loss function can be back-propagated. Therefore, the conventional unit step function is not applicable, but the following approximate unit step function is used, namely:

[0143]

[0144] Among them, γ is the scaling parameter.

[0145] In addition, when the scaling parameter γ is a large non-negative real number, the effect of the function is very close to the unit step function and is smooth and differentiable. Therefore, during training, the approximate unit step function can help the system achieve adaptive optimization of the threshold t i In order to simulate the actual application, the conventional unit step function is used as the quantization function during the test, and the threshold t i Fixed to the optimal threshold obtained during the training process.

[0146] Specifically, the quantized signal q obtained after two-bit quantization in the embodiment of the present application includes the true coded signal e r Therefore, it can be used as conditional information and input into the generator together with the random signal to obtain the approximate true coding signal e r The pseudo-coded signal e f , to achieve the goal of channel equalization, the generator contains multiple layers of one-dimensional convolutional layers to learn the impact of the channel on the transmitted signal. Then, in order to simulate the coded signal e r , several fully connected layers are added after the one-dimensional convolution layer, and normalized at the end. The pseudo-coded signal e is obtained after generator equalization f It will be input into the decoding network for decoding, which is more conducive to the symbol s i reconstruction.

[0147] In step S33, the embodiment of the present application can fuse the decoding network of the autoencoder and the discriminator of the generative adversarial network into one network, which can not only judge the authenticity of the encoded signal to compete with the generator, but also decode the encoded signal and restore it to the input signal.

[0148] Among them, the discriminator can convert the true encoded signal e generated by the encoding network r The pseudo-coded signal e generated by the generator f The decoding network aims to recover the symbol s from the encoded signal. i , it can be seen that the discriminator and the decoding network can input the same information, and at the same time lay the foundation for the discriminator and the decoding network to share the same network.

[0149] Furthermore, from the output of the discriminator and the decoding network, the discriminator solves a binary classification problem, while the decoding network solves an M-classification problem, where each category represents a symbol s i In order to enable the discriminator and the decoding network to share the same network DD to reduce the overall size of the system and avoid mutual interference caused by different output targets of the two, the output dimension of DD becomes M+1, and the first M elements are responsible for M classification tasks to recover the symbol s i , the last element is used to judge the true or false of the coded signal. When the coded signal is true, the coded signal e rWhen , the value of the last element is expected to be close to 1. When the coded signal is a false coded signal e f , then we expect the last element to have a value close to 0.

[0150] In addition, in order to enable DD to not only distinguish true and false coded signals, but also to determine whether the coded signal corresponds to the correct quantized signal, the quantized signal q is also used as part of the shared network input. At the same time, the quantized signal q also contains the symbol s i information, which is helpful for symbol s i reconstruction.

[0151] Step S34, the embodiment of the present application is based on a weighted loss function and a training strategy with turning points, which ensures that the end-to-end training of the autoencoder and the game training of the generative adversarial network can be carried out simultaneously, and the performance of the trained system is tested.

[0152] In summary, DD can not only determine the authenticity of the coded signal, but also reconstruct it into the transmitted symbol. Therefore, the loss function is the weighted sum of the two objectives, namely:

[0153]

[0154] Among them, α is the weight coefficient that weighs the discrimination ability and symbol reconstruction accuracy, The goal of distinguishing true from false is as shown below.

[0155] Distinguishing true from false goals The expression is:

[0156]

[0157] Among them, DD b (·) gives the probability that DD judges the coded signal as a true coded signal.

[0158] and Recover the transmitted symbol s from the true and false coded signals respectively i The goal is expressed as:

[0159]

[0160]

[0161] Among them, DD m (·) is the reconstruction of the transmitted symbol.

[0162] For the generator, it is possible to generate sufficiently realistic encoded signals to confuse DD. The goal is expressed as:

[0163]

[0164] Furthermore, the generator can determine which type of symbol s the pseudo-encoded signal should be reconstructed into by DD i , so it can also be obtained by Description, the loss function of the generator is also the weighted sum of the above two parts, which can be expressed as:

[0165]

[0166] Among them, β is the weight coefficient.

[0167] During training, the symbol s sent i First, the true coded signal e is obtained through the coding network r Then, the pseudo-coded signal e is obtained by 16QAM modulation, wireless Rice channel, two-bit quantization and generator. f , DD is used to distinguish true and false coded signals and restore them to symbols s according to the quantized signal q i During the training process, the parameters of the encoding network and the discriminator parameters are optimized as a group, so that the end-to-end joint optimization of the autoencoder can be achieved, while the generator parameters are another group, and the two groups of parameters are trained alternately and cyclically, thus ensuring the game optimization of the generative adversarial network.

[0168] In addition, the training process contains a key turning point, before which more emphasis is placed on training the accuracy of symbol recovery, i.e. and The weight is higher, It can make the communication system obtain an optimal standard based on the ideal channel. It can prevent overfitting. After the turning point, the parameters of the coding network will be fixed, so that the generator and DD have fixed conditional information, that is, the quantized signal q, which is more conducive to the stability of training. After the turning point, the weights become 0.5, so that the fitting ability of the generator is gradually improved, and the gradual approach of the true and false coding signals will further improve the accuracy of DD in recovering the sent symbols.

[0169] The structure of the test process is simpler than that of the training process, because there is no need for DD to distinguish true and false encoded signals. The true encoded signal from the encoder is r It will not directly enter DD, but will only go through 16QAM modulation, Rice channel and two-bit quantization in sequence to obtain the quantized signal q. The trained generator will obtain the approximate true coded signal e based on the quantized signal q. r The pseudo-coded signal e f , and enter DD for symbol s i recovery. Specific embodiment 4

[0171] For example, the embodiment of the present application may use OFDM modulation to modulate the true coded signal, wherein the channel is a free space optical channel, and eight-bit quantization is used to reduce energy consumption. The method includes the following steps:

[0172] In step S41, the embodiment of the present application encodes the input signal through the encoding network of the automatic encoder, and the obtained encoded signal is called a true encoded signal.

[0173] Specifically, the embodiment of the present application can treat each k bits of information in the bit sequence as a group to form a symbol s i , then the number of symbol types is M = 2 k , the input of the encoding network is s i The one-hot representation is that the i-th value of the M-dimensional zero vector is set to 1, where the encoding network consists of several fully connected layers to convert the symbol s i Mapping to coded signal And normalize it at the end. Taking the j-th fully connected layer as an example, the process can be expressed as:

[0174] h j =f ReLU (W j h j-1 +b j ),

[0175] Among them, h j , W j 、b j are the feature vector, weight matrix and offset of the jth layer, respectively, ReLU An activation function for nonlinear fitting.

[0176] In addition, since the uncoded bit stream information has poor anti-interference ability, the coding network can improve the reliability of transmission by increasing redundancy.

[0177] In step S42, the embodiment of the present application can send the true coded signal after OFDM modulation, and reach the receiving end through the free space optical channel. The receiving end performs eight-bit quantization to obtain a quantized signal, and uses the quantized signal as conditional information to fit the random signal into a false coded signal that approximates the true coded signal through the generator of the generative adversarial network.

[0178] Furthermore, eight-bit quantization at the receiving end can reduce power consumption. The quantization value is obtained by dividing the received signal y by the threshold t i (i=1,2,…,8) is used for comparison and decision. However, due to channel interference, the optimal threshold cannot be determined. Therefore, it can be used as a trainable parameter. During the training process, it is adaptively adjusted to the optimal threshold that can minimize the system bit error rate, and the threshold t is optimized. i, and the quantization function is smooth and differentiable, so that the gradient of the loss function can be back-propagated. Therefore, the conventional unit step function is not applicable, but the following approximate unit step function is used, namely:

[0179]

[0180] Among them, γ is the scaling parameter.

[0181] In addition, when the scaling parameter γ is a large non-negative real number, the effect of the function is very close to the unit step function and is smooth and differentiable. Therefore, during training, the approximate unit step function can help the system achieve adaptive optimization of the threshold t i In order to simulate the actual application, the conventional unit step function is used as the quantization function during the test, and the threshold t i Fixed to the optimal threshold obtained during the training process.

[0182] Furthermore, the quantized signal q obtained after eight-bit quantization in the embodiment of the present application includes the true coded signal e r Therefore, it can be used as conditional information and input into the generator together with the random signal to obtain the approximate true coding signal e r The pseudo-coded signal e f , to achieve the goal of channel equalization, the generator contains multiple layers of one-dimensional convolutional layers to learn the impact of the channel on the transmitted signal. Then, in order to simulate the coded signal e r , several fully connected layers are added after the one-dimensional convolution layer, and normalized at the end. The pseudo-coded signal e obtained after generator equalization f It will be input into the decoding network for decoding, which is more conducive to the symbol s i reconstruction.

[0183] Step S43, the embodiment of the present application can merge the decoding network of the autoencoder and the discriminator of the generative adversarial network into one network, which can not only judge the authenticity of the encoded signal to compete with the generator, but also decode the encoded signal and restore it to the input signal.

[0184] Among them, the role of the discriminator is to convert the true encoding signal e generated by the encoding network r The pseudo-coded signal e generated by the generator f The decoding network aims to recover the symbol s from the encoded signal. i , it can be seen that the discriminator and the decoding network can input the same information, and at the same time lay the foundation for the discriminator and the decoding network to share the same network.

[0185] Furthermore, from the output of the discriminator and the decoding network, the discriminator solves a binary classification problem, while the decoding network solves an M-classification problem, where each category represents a symbol si In order to enable the discriminator and the decoding network to share the same network DD to reduce the overall size of the system and avoid mutual interference caused by different output targets of the two, the output dimension of DD becomes M+1, and the first M elements are responsible for M classification tasks to recover the symbol s i , the last element is used to judge the true or false of the coded signal. When the coded signal is true, the coded signal e r When , the value of the last element is expected to be close to 1. When the coded signal is a false coded signal e f , then we expect the last element to have a value close to 0.

[0186] In addition, in order to enable DD to not only distinguish true and false coded signals, but also to determine whether the coded signal corresponds to the correct quantized signal, the quantized signal q is also used as part of the shared network input. At the same time, the quantized signal q also contains the symbol s i information, which is helpful for symbol s i reconstruction.

[0187] In step S44, the embodiment of the present application is based on a weighted loss function and a training strategy with turning points, which ensures that the end-to-end training of the autoencoder and the game training of the generative adversarial network can be carried out simultaneously, and a performance test is performed on the trained system.

[0188] In summary, DD can not only determine the authenticity of the coded signal, but also reconstruct it into the transmitted symbol. Therefore, the loss function is the weighted sum of the two objectives, namely:

[0189]

[0190] Among them, α is the weight coefficient that weighs the discrimination ability and symbol reconstruction accuracy, The goal of distinguishing true from false is as shown below.

[0191] Distinguishing true from false goals The expression is:

[0192]

[0193] Among them, DD b (·) gives the probability that DD judges the coded signal as a true coded signal.

[0194] and Recover the transmitted symbol s from the true and false coded signals respectively i The goal is expressed as:

[0195]

[0196]

[0197] Among them, DD m (·) is the reconstruction of the transmitted symbol.

[0198] For the generator, it is possible to generate sufficiently realistic encoded signals to confuse DD. The goal is expressed as:

[0199]

[0200] Furthermore, the generator can determine which type of symbol s the pseudo-encoded signal should be reconstructed into by DD i , so it can also be obtained by Description, the loss function of the generator is also the weighted sum of the above two parts, which can be expressed as:

[0201]

[0202] Among them, β is the weight coefficient.

[0203] During training, the symbol s sent i First, the true coded signal e is obtained through the coding network r Then, it goes through OFDM modulation, free space optical channel, eight-bit quantization and generator to obtain the pseudo-coded signal e f , DD is used to distinguish true and false coded signals and restore them to symbols s according to the quantized signal q i During the training process, the parameters of the encoding network and the discriminator parameters are optimized as a group, so that the end-to-end joint optimization of the autoencoder can be achieved, while the generator parameters are another group, and the two groups of parameters are trained alternately and cyclically, thus ensuring the game optimization of the generative adversarial network.

[0204] In addition, the training process contains a key turning point, before which more emphasis is placed on training the accuracy of symbol recovery, i.e. and The weight is higher, It can make the communication system obtain an optimal standard based on the ideal channel. It can prevent overfitting. After the turning point, the parameters of the coding network will be fixed, so that the generator and DD have fixed conditional information, that is, the quantization signal q, which is more conducive to the stability of training. After the turning point, the weights become 0.5, so that the fitting ability of the generator is gradually improved, and the gradual approach of the true and false coding signals will further improve the accuracy of DD in recovering the sent symbols.

[0205] Furthermore, the structure of the test process is simpler than that of the training process, because there is no need for DD to distinguish true and false encoded signals. The true encoded signal from the encoder is rIt will not directly enter DD, but will only go through OFDM modulation, free space optical channel and eight-bit quantization in sequence to obtain the quantized signal q. The trained generator will obtain the approximate true coded signal e based on the quantized signal q. r The pseudo-coded signal e f , and enter DD for symbol s i recovery.

[0206] According to the communication method based on the fusion of the autoencoder and the generative adversarial network proposed in the embodiment of the present application, the input signal can be encoded by the encoding network of the autoencoder to obtain a true coded signal and modulate it, and the received signal after passing through the channel is quantized to obtain a quantized signal, so that the random signal is fitted to a false coded signal that approximates the true coded signal by the generator of the generative adversarial network, and the decoding network of the autoencoder and the discriminator of the generative adversarial network are fused into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, and a fusion network that judges the true or false of the coded signal to play a game with the generator and decodes any coded signal to restore it to a transmission signal is obtained, so that accurate channel information can be obtained in a communication system with a complex channel, and the reliability and accuracy of the communication system are effectively improved. Thus, the problem that the communication in the related art is more dependent on the channel model, resulting in the inability to obtain accurate channel information in a communication system with a complex channel, increasing the complexity of system implementation, and reducing the reliability and accuracy of the communication system is solved.

[0207] Next, a communication system based on a fusion autoencoder and a generative adversarial network proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0208] Figure 5 It is a block diagram of a communication system based on a fusion autoencoder and a generative adversarial network according to an embodiment of the present application.

[0209] like Figure 5 As shown, the communication system 10 based on the fusion of the autoencoder and the generative adversarial network includes: an encoding module 100, a first processing module 200, a fitting module 300 and a second processing module 400.

[0210] Specifically, the encoding module 100 is used to encode the input signal through the encoding network of the automatic encoder to obtain a true encoded signal.

[0211] The first processing module 200 is used to modulate the true coded signal and quantize the received signal after passing through the channel to obtain a quantized signal.

[0212] The fitting module 300 is used to fit the random signal into a false coding signal that approximates the true coding signal based on the quantized signal through the generator of the generative adversarial network.

[0213] The second processing module 400 is used to fuse the decoding network of the autoencoder and the discriminator of the generative adversarial network into an initial network, and train the initial network based on a preset weighted loss function and a training strategy with a turning point, so as to obtain a fused network that judges the truth or falsity of the encoded signal to compete with the generator and decodes any encoded signal to restore it to a transmission signal.

[0214] Optionally, in one embodiment of the present application, the encoding module 100 includes: a generating unit and a processing unit.

[0215] The generating unit is used to generate corresponding symbols according to each preset bit information of the bit sequence, and map the symbols into corresponding coding information.

[0216] The processing unit is used to perform normalization processing based on the coding information to obtain a true coding signal.

[0217] Optionally, in one embodiment of the present application, the first processing module 200 includes: a modulation unit and a sending unit.

[0218] Among them, the modulation unit is used to modulate the true coded signal based on binary amplitude keying OOK, two-phase phase shift keying BPSK, orthogonal amplitude modulation QAM or orthogonal frequency division multiplexing modulation OFDM modulation.

[0219] The sending unit is used to send the modulated true coded signal using the sending end.

[0220] Optionally, in one embodiment of the present application, the first processing module 200 is further used to use the quantization function quantization threshold as a trainable parameter based on an adaptive quantization method to adaptively adjust to an optimal threshold.

[0221] Optionally, in one embodiment of the present application, the quantization function is an approximate unit step function.

[0222] Optionally, in one embodiment of the present application, the training strategy with a turning point includes that before the turning point, the loss function weight of the autoencoder is greater than a preset threshold, and after the turning point, the parameters of the encoding network are fixed, and the loss function weights of the autoencoder and the generative adversarial network are the same.

[0223] It should be noted that the aforementioned explanation of the embodiment of the communication method based on the fusion of the autoencoder and the generative adversarial network is also applicable to the communication system based on the fusion of the autoencoder and the generative adversarial network of this embodiment, and will not be repeated here.

[0224] According to the communication system based on the fusion of the autoencoder and the generative adversarial network proposed in the embodiment of the present application, the input signal can be encoded by the encoding network of the autoencoder to obtain a true coded signal and modulate it, and the received signal after passing through the channel is quantized to obtain a quantized signal, so that the random signal is fitted to a false coded signal that approximates the true coded signal by the generator of the generative adversarial network, and the decoding network of the autoencoder and the discriminator of the generative adversarial network are fused into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, and a fusion network that judges the true or false of the coded signal to play a game with the generator and decodes any coded signal to restore it to a transmission signal is obtained, so that accurate channel information can be obtained in a communication system with a complex channel, and the reliability and accuracy of the communication system are effectively improved. Thus, the problem that the communication in the related art is more dependent on the channel model, resulting in the inability to obtain accurate channel information in a communication system with a complex channel, increasing the complexity of system implementation, and reducing the reliability and accuracy of the communication system is solved.

[0225] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0226] A memory 601 , a processor 602 , and a computer program stored in the memory 601 and executable on the processor 602 .

[0227] When the processor 602 executes the program, the communication method based on the fusion of the autoencoder and the generative adversarial network provided in the above embodiment is implemented.

[0228] Furthermore, the electronic device further comprises:

[0229] The communication interface 603 is used for communication between the memory 601 and the processor 602 .

[0230] The memory 601 is used to store computer programs that can be executed on the processor 602 .

[0231] The memory 601 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0232] If the memory 601, the processor 602 and the communication interface 603 are implemented independently, the communication interface 603, the memory 601 and the processor 602 can be connected to each other through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0233] Optionally, in a specific implementation, if the memory 601, the processor 602 and the communication interface 603 are integrated on a chip, the memory 601, the processor 602 and the communication interface 603 can communicate with each other through an internal interface.

[0234] The processor 602 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0235] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned communication method based on a fusion autoencoder and a generative adversarial network.

[0236] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0237] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0238] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0239] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or N wirings (electronic devices), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways as necessary and then storing it in a computer memory.

[0240] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0241] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0242] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0243] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A communication method based on fusion autoencoder and generative adversarial network, characterized in that: The following steps are involved: Encode the input signal through the encoding network of the autoencoder to obtain the true encoded signal; Modulating the true coded signal, and quantizing the received signal after passing through the channel to obtain a quantized signal; Based on the quantized signal, a random signal is fitted into a false coding signal that approximates a true coding signal through a generator of a generative adversarial network; as well as The decoding network of the autoencoder and the discriminator of the generative adversarial network are fused into an initial network, and the initial network is trained based on a preset weighted loss function and a training strategy with a turning point, so as to obtain a fusion network that judges the truth or falsity of a coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal, wherein the preset weighted loss function is a weighted sum of the loss function of a conventional generative adversarial network and the loss function of the autoencoder; the training strategy with a turning point includes: before the turning point, the weight of the loss function of the autoencoder is greater than a preset threshold, and after the turning point, the parameters of the coding network are fixed, and the weights of the loss functions of the autoencoder and the generative adversarial network are the same.

2. The method according to claim 1, characterized in that: The step of encoding the input signal through the encoding network of the automatic encoder to obtain a true encoded signal includes: Generate a corresponding symbol according to each preset bit information of the bit sequence, and map the symbol into corresponding coded information; Normalization processing is performed based on the coding information to obtain the true coding signal.

3. The method according to claim 1, characterized in that The step of modulating the true coded signal comprises: Modulating the true coded signal based on binary amplitude keying (OOK), binary phase shift keying (BPSK), quadrature amplitude modulation (QAM), or orthogonal frequency division multiplexing (OFDM) modulation; The modulated true coded signal is sent by the transmitter.

4. The method according to claim 1 or 3, characterized in that: The step of quantizing the received signal after passing through the channel to obtain a quantized signal includes: Based on the adaptive quantization method, the quantization threshold of the quantization function is used as a trainable parameter to adaptively adjust to the optimal threshold.

5. The method according to claim 4, characterized in that The quantization function is an approximate unit step function.

6. A communication system based on a fusion autoencoder and a generative adversarial network, characterized in that: include: An encoding module, used for encoding the input signal through an encoding network of the automatic encoder to obtain a true encoded signal; A first processing module, used for modulating the true coded signal, and quantizing the received signal after passing through the channel to obtain a quantized signal; A fitting module, configured to fit a random signal into a false coding signal approximating a true coding signal through a generator of a generative adversarial network based on the quantized signal; as well as The second processing module is used to fuse the decoding network of the autoencoder and the discriminator of the generative adversarial network into an initial network, and train the initial network based on a preset weighted loss function and a training strategy with a turning point, so as to obtain a fused network that judges the truth or falsity of a coded signal to compete with the generator and decodes any coded signal to restore it to a transmission signal, wherein the preset weighted loss function is a weighted sum of the loss function of a conventional generative adversarial network and the loss function of the autoencoder; the training strategy with a turning point includes: before the turning point, the weight of the loss function of the autoencoder is greater than a preset threshold, and after the turning point, the parameters of the coding network are fixed, and the loss function weights of the autoencoder and the generative adversarial network are the same.

7. The system according to claim 6, characterized in that The encoding module comprises: A generating unit, configured to generate a corresponding symbol according to each preset bit information of the bit sequence, and map the symbol into corresponding coding information; A processing unit is used to perform normalization processing based on the coding information to obtain the true coding signal.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the communication method based on a fusion autoencoder and a generative adversarial network as described in any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the communication method based on the fusion of the autoencoder and the generative adversarial network as described in any one of claims 1 to 5.

Images