A Smart Transceiver Design Method Based on RCNN-GAN

Abstract

This invention relates to the field of wireless communication technology, specifically to a method for designing an intelligent transceiver. The method includes steps such as intelligent transceiver model construction, acquisition of training and test sets, training and testing, and obtaining the intelligent transceiver. It utilizes an attention-based residual-assisted generative adversarial network (AAR-GAN) model and a residual autoencoder (RCNN) model, which respectively possess coding / modulation, channel, demodulation, and decoding functions. Compared to CNN-CGAN, these two neural networks not only improve the quality of generated channel effects but also comprehensively enhance BER performance and strengthen the stability of the network models.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically to a design method for an intelligent transceiver. Background Technology

[0002] In traditional wireless communication systems, data transmission requires processing by multiple signal processing modules in both the transmitter and receiver. Therefore, designing and optimizing individual modules to achieve global system optimization is a common approach. Furthermore, channel propagation is an assumed mathematical model embedded in the design, which may not accurately reflect actual transmission scenarios. To address the shortcomings of this assumed channel model, applying deep learning to traditional communication systems to jointly optimize traditional modules, such as channel detection and channel estimation, is also a common practice. However, these two methods cannot perform joint optimization during the training process, as this increases the complexity of the system design and raises the development cost of the program.

[0003] To address the aforementioned issues, existing technologies often employ deep neural networks (DNNs) to jointly optimize transceivers. However, these methods are implemented under known channel models, and since real-world channel information is unknown, their actual optimization effectiveness is limited. Consequently, channel simulation-based solutions are frequently used, utilizing deep generative models to simulate the output distribution of real channels.

[0004] Existing technologies employ Generative Adversarial Networks (GANs) to simulate real channel outputs. However, the signals generated by GANs are unstable, resulting in mediocre simulations of real channel outputs. To address this issue, researchers have proposed a data-driven approach using Conditional GANs (CGANs) to capture channel effects, incorporating received pilot signals as part of the CGAN's conditional information to simulate real channel outputs (referred to as the CNN-CGAN scheme). However, this approach still suffers from poor network model stability and low bit error rate (BER) performance. Summary of the Invention

[0005] The purpose of this invention is to provide a smart transceiver design method based on RCNN-GAN, which has better channel simulation generation capabilities and stronger stability, thereby comprehensively improving BER performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for designing an intelligent transceiver based on RCNN-GAN, characterized by the following steps:

[0008] S01, Steps for constructing a smart transceiver model;

[0009] Attention-based residual-assisted generative adversarial network (AAR-GAN) model to simulate the channel;

[0010] The transmitter and receiver are built based on the residual autoencoder RCNN model;

[0011] S02, Steps for obtaining training and test sets;

[0012] Randomly generate information bit stream s; randomly sample instantaneous channel state information CSI and noise z from the channel dataset; obtain the training set A1 and test set A2 of the AAR-GAN model and the training set B1 and test set B2 of the RCNN model;

[0013] S03, Training Steps;

[0014] The AAR-GAN model is trained using the training set A1, the parameters of the AAR-GAN model are updated, and the AAR-GAN model is iteratively trained until convergence.

[0015] The RCNN model is trained using training set B1, the parameters of the RCNN model are updated, and the RCNN model is trained iteratively until convergence.

[0016] S04, Testing and Smart Transceiver Acquisition Steps;

[0017] Input the test set A2 and the test set B2 into the converged AAR-GAN model and RCNN model to obtain the bit error rate (BER). If the BER is higher than the design standard value, continue iterating. If the BER is lower than the design standard value, end the iteration and obtain the intelligent transceiver.

[0018] The attention-based residual-assisted generative adversarial network (AAR-GAN) model consists of two network modules: a residual generator and a discriminator. The residual generator module comprises a three-layer network and a channel attention mechanism, SE-Block. The discriminator module also comprises a three-layer network: the first and second layers are both convolutional layers with ReLU activation; the third layer is a convolutional layer with sigmoid activation, which is used to simulate a channel.

[0019] The residual autoencoder RCNN model consists of an encoder and a decoder. The encoder acts as the transmitter, and the decoder acts as the receiver.

[0020] The TensorFlow system randomly generates an information bitstream s of length K, and randomly samples instantaneous channel state information (CSI) and noise z following the N(0,1) order from AWGN, Rayleigh, and DeepMIMO channel datasets. The information bitstream s is transmitted through a transmitter to obtain the transmitted signal x, the instantaneous channel state information CSI, and the noise z. These three components constitute the training set A1 and test set A2 of the AAR-GAN model and the training set B1 and test set B2 of the RCNN model.

[0021] Using training set A1 and test set A2, a binary cross-entropy function and an L2 regularization term are used as loss functions. The Adam optimization algorithm is then used to train the generator and discriminator of the AAR-GAN model, updating the AAR-GAN network model parameters. Simultaneously, using training set B1 and test set B2, a binary cross-entropy function is used as the loss function, and the Adam optimization algorithm is used to train the R-CNN model, updating the R-CNN model parameters. The AAR-GAN and R-CNN network models are trained iteratively until they converge (i.e., the loss functions decrease and tend to stabilize).

[0022] After each round of training iterations, testing begins. Test sets A2 and B2 are input into the converged AAR-GAN and RCNN models to obtain the Bit Error Rate (BER). The learned intelligent transceiver model is then tested for BER performance on a real channel. If the BER is higher than the design standard, iteration continues; if the BER is lower than the design standard, iteration ends, and the intelligent transceiver is obtained. The BER is calculated by dividing the number of erroneous bits over a period of time by the total number of bits transmitted, and is typically expressed as a percentage.

[0023] In summary, the intelligent transceiver designed in this invention possesses all the modules of a traditional transceiver, all of which are composed of neural networks. For the channel model, an AAR-GAN model based on an attention mechanism is designed, and channel effects are simulated on three channel datasets: AWGN, Rayleigh, and DeepMIMO. This improves the channel simulation generation capability and stability, and comprehensively enhances BER performance.

[0024] As a preferred embodiment of the present invention, in S01, the AAR-GAN model includes a residual generator and a discriminator; the residual generator includes three convolutional layer networks and a channel attention mechanism SE-Block network; the discriminator includes two convolutional layers and a ReLU activation layer network and a convolutional layer and a sigmoid activation layer network.

[0025] The residual generator consists of three convolutional layers and a channel attention mechanism, SE-Block: the first layer is a convolutional layer (output feature T1), SE-Block, batch normalization (BN) layer, and ReLU activation layer; the second layer is a convolutional layer, BN layer, and ReLU activation layer; the third layer is a convolutional layer (output feature T2), and finally T2 is added to the emitted signal x, as shown in the following formula.

[0026]

[0027] Where G(z,v) and G is the output of the residual generator. r (z,v) represents the transmitted signal and the output signal. The difference is that z represents the noise from random sampling, and v represents the conditional information, i.e., the transmitted signal x and the pilot signal y. p .

[0028] The SE-Block module primarily consists of compression (Squeeze) and excitation operations. The compression operation comprises a single layer: a global average pooling layer. The excitation operation comprises two layers: the first layer consists of a convolutional layer and a ReLU activation layer; the second layer consists of a convolutional layer and a sigmoid activation layer. Finally, a scaling operation is used to multiply and weight the output of the excitation operation onto the previous features, completing the recalibration of the original features along the channel dimension.

[0029] The formula for global average pooling in the compression operation is as follows:

[0030]

[0031] Where B and K are the width and height of T1, respectively, T1 is the output of the first convolutional layer, and "×" represents the multiplication operation.

[0032] The formula for the stimulus operation is as follows:

[0033]

[0034] In this context, "*" represents a convolution operation, W1 and W2 are weights, and b1 and b2 are biases.

[0035] The formula for the scale operation is as follows:

[0036] T2 = T1 × O2 (4)

[0037] T2 is the output of the SE-Block module.

[0038] The discriminator consists of three layers: the first and second layers are both convolutional layers and ReLU activation layers; the third layer is a convolutional layer and a sigmoid activation layer.

[0039] As a preferred embodiment of the present invention, in S01, the RCNN model includes an encoder and a decoder, wherein the encoder constitutes a transmitter and the decoder constitutes a receiver.

[0040] The encoder consists of four layers: the first to third layers are all convolutional layers and ReLU activation layers; the fourth layer is a convolutional layer.

[0041] The decoder consists of eight layers: layers one through seven are convolutional layers and ReLU activation layers; the eighth layer is a convolutional layer and a sigmoid activation layer, with skip connections between the second convolutional layer and the fourth ReLU layer, and between the fifth convolutional layer and the seventh ReLU layer. The sigmoid layer ensures that all elements in the output vector are within the range [0,1], thus achieving multi-label classification.

[0042] The transmitter is composed of an RCNN encoder, and the receiver is composed of a decoder.

[0043] As a preferred embodiment of the present invention, in step S02, the channel dataset includes the AWGN channel dataset, the Rayleigh channel dataset, and the DeepMIMO channel dataset.

[0044] Because the propagation path between the transmitter and receiver in wireless communication systems is very complex and extremely time-varying and random, channel models are particularly difficult to analyze. Therefore, it is necessary to combine reality with channel simulation and collect channel datasets by selecting multiple channel types.

[0045] An AWGN channel is a common idealized channel model. Its instantaneous values ​​follow a Gaussian distribution, and its power spectral density follows a uniform distribution. Additive white Gaussian noise plays an important role in both theoretical and practical applications, and is therefore frequently used in communication system simulations.

[0046] The Rayleigh fading channel is a statistical model of the radio signal propagation environment. This model assumes that the signal amplitude, or "fading," occurs randomly after passing through the wireless channel, and that its envelope follows a Rayleigh distribution. This channel model can describe shortwave channels reflected by the ionosphere and troposphere, as well as densely built-up urban environments.

[0047] Deep MIMO (Multiple-Input Multiple-Output) refers to a technology in wireless communication that uses multiple antennas to transmit and receive signals. In a MIMO system, both the transmitting and receiving ends are equipped with multiple antennas. By transmitting signals between these multiple antennas, data transmission rate and reliability can be improved.

[0048] For each of the three channels mentioned above, corresponding datasets can be obtained, improving the model's ability to effectively capture the complex characteristics of real channels and reducing the impact of different channel models on communication quality.

[0049] As a preferred embodiment of the present invention, for AWGN channels, there is no need to sample instantaneous channel state information (CSI); for Rayleigh channels, the sampled instantaneous channel state information (CSI) follows a complex Gaussian distribution with a mean of 0 and a variance of 1; for DeepMIMO channels, the sampled instantaneous channel state information (CSI) is based on a ray-traced channel dataset.

[0050] As a preferred embodiment of the present invention, in step S03, during the training process of the AAR-GAN model, the binary cross-entropy function and L2 regularization are used as loss functions.

[0051] The formulas for the binary cross-entropy function and L2 regularization as the loss function in the AAR-GAN model are as follows:

[0052]

[0053] Among them, L G The loss function of the residual generator. L represents the L2 regularization term, L D Let θ be the loss function of the discriminator. G and θ D Let z represent the parameters of the generator and discriminator, respectively, and z represent the sampled from sample p. z The random noise, y represents the random noise sampled from the real data sample p. d The distribution of λ is the weight decay parameter.

[0054] Let the loss function L of the residual generator be... G The loss function L of the discriminator D The AAR-GAN model converges when it decreases to a minimum and tends to stabilize.

[0055] As a preferred embodiment of the present invention, in step S03, the binary cross-entropy function is used as the loss function during the training process of the RCNN model.

[0056] The formula for the binary cross-entropy function as the loss function in the RCNN model is as follows:

[0057]

[0058] in, Let s represent the binary cross-entropy loss function. i p is the i-th element of s i For s i The probability obtained after sigmoid activation.

[0059] Let the binary cross-entropy loss function The RCNN model converges when it decreases to a minimum and tends to stabilize.

[0060] As a preferred embodiment of the present invention, in step S03, the Adam algorithm is used to optimize the AAR-GAN model and the RCNN model.

[0061] In summary, the present invention has the following beneficial effects:

[0062] The method of this invention comprises two neural network models: an attention-based residual-assisted generative adversarial network (AAR-GAN) model and a residual autoencoder (RCNN) model, which respectively possess encoding / modulation, channel, demodulation, and decoding functions. Compared to CNN-CGAN, these two neural networks not only improve the quality of the generated channel effects but also enhance BER performance.

[0063] Meanwhile, channel effects were simulated on three channel datasets: AWGN, Rayleigh, and DeepMIMO, improving communication quality. Furthermore, considering the network degradation problem that occurs with increasing network depth, skip connections were added to the RCNN decoder, enabling the model to develop deeper and more powerful neural networks, achieving better intelligent transceiver performance. Attached Figure Description

[0064] Figure 1 This is a flowchart of a smart transceiver design method based on RCNN-GAN;

[0065] Figure 2 This is a model diagram of an intelligent transceiver design method based on RCNN-GAN;

[0066] Figure 3 This is a schematic diagram of the structure of the AAR-GAN model of this invention;

[0067] Figure 4 This is a schematic diagram of the structure of the RCNN model of the present invention;

[0068] Figure 5This is a comparison of the BER (Bit Rate) of an intelligent transceiver based on RCNN-GAN (RCNN-AAR-GAN) and an intelligent transceiver based on deep learning (CNN-CGAN) under an AWGN channel.

[0069] Figure 6 This is a comparison of the BER (BER) of an intelligent transceiver based on RCNN-GAN (RCNN-AAR-GAN) and an intelligent transceiver based on deep learning (CNN-CGAN) in the Rayleigh channel.

[0070] Figure 7 This is a comparison of the BER (Bitrate Error) of an intelligent transceiver based on RCNN-GAN (RCNN-AAR-GAN) and an intelligent transceiver based on deep learning (CNN-CGAN) under a DeepMIMO channel. Detailed Implementation

[0071] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. Any person may implement the present disclosure in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0072] The term "comprising" and its variations as used herein signify open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "at least partially based on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment".

[0073] like Figure 1As shown, an intelligent transceiver model based on R-CNN and AAR-GAN models is constructed using TensorFlow. Then, signal bits are randomly generated, instantaneous channel state information (CSI) is sampled from the channel set, and noise is randomly sampled to obtain training sets A1 and A2 for the AAR-GAN model and training sets B1 and B2 for the R-CNN model. In each training round, training set A1 is used, with binary cross-entropy and L2 regularization as loss functions, and the Adam optimization algorithm is used to train the AAR-GAN network model to update its parameters. Next, training set B1 is used, with binary cross-entropy as the loss function, and the Adam optimization algorithm is used to train the R-CNN model to update its parameters. Then, in the testing phase, the BER performance of the learned intelligent transceiver is tested on a real channel using test sets A2 and B2. Training and testing are iterated multiple times until the BER no longer decreases during testing, at which point training and testing are stopped, thus completing the entire intelligent transceiver design.

[0074] like Figure 2 As shown, the AAR-GAN model consists of two parts: a residual generator and a discriminator.

[0075] The first step involves concatenating the emission signal x, conditional v, and random noise z from the RCNN model to obtain z. c The input is fed into the first layer of the residual generator in the AAR-GAN model. The first layer is a one-dimensional convolutional layer with 64 filters and a kernel size of 3, resulting in a feature vector T1. T1 is then input into SE-Block, where it undergoes global average pooling to obtain a feature vector G of size (1,1,C). p The first feature vector is passed through a convolutional layer with a kernel size of 3 and the ReLU activation function. Then, it passes through a second convolutional layer with a kernel size of 3 and the sigmoid activation function. The feature vector is then restored to its original dimensions, resulting in feature vector T2. T2 is then passed through a BN layer and a ReLU activation layer before being input into the second layer, which includes a convolutional layer with 64 filters and a kernel size of 3, a BN layer, and a ReLU activation layer. This is then input into the third layer, which contains a one-dimensional convolutional layer with 2 filters and a kernel size of 3, and no activation layer. The output is [output value missing]. Finally, O[value missing]. g The signal is added to the transmitted signal x to form a skip connection, resulting in the generated received signal. In this way, the residual generator learns the difference between the input and the output.

[0076] The second step involves inputting the transmitted signal x into the actual channel to obtain the received signal y, and then concatenating y and v to obtain y'.c , By concatenating with v Then y c and The inputs are fed into the discriminator, which consists of three layers. The first and second layers are both convolutional layers and ReLU activation layers, with 64 and 3 filters in the convolutional layers. The third layer is a convolutional layer and a sigmoid activation layer, with 1 and 3 filters in the convolutional layers. The sigmoid layer ensures that all elements in the output vector are within the range [0,1], thus implementing a multi-label classification method. The discriminator's role is to determine the value of y. c and The difference is used to better assist the residual generator in generating samples that are closer to real data in order to deceive the discriminator.

[0077] like Figure 4 As shown, the RCNN model includes an encoder and a decoder, where the encoder performs the function of a transmitter and the decoder performs the function of a receiver. The first step involves inputting the information bitstream s into the encoder. The first layer consists of a one-dimensional convolutional layer and a ReLU activation layer, with 256 filters and a filter size of 5 in the convolution. The second layer consists of a one-dimensional convolutional layer and a ReLU activation layer, with 128 filters and a filter size of 3 in the convolution. The third layer consists of a one-dimensional convolutional layer and a ReLU activation layer, with 64 filters and a filter size of 3 in the convolution. The fourth layer consists of a one-dimensional convolutional layer and a ReLU activation layer, with 2 filters and a filter size of 3 in the convolution. The second step involves passing the transmitted signal x through a channel to obtain the received signal y, which is then input into the decoder. This decoder consists of eight layers. The first layer comprises a one-dimensional convolutional layer and a ReLU activation layer, with 256 filters and a size of 5. Layers two through four are also convolutional layers and ReLU activation layers, with 128 filters and a size of 5. Layers five and six are one-dimensional convolutional layers with 64 filters and a kernel size of 5. Layer seven is a one-dimensional convolutional layer with 64 filters and a kernel size of 3. Skip connections are introduced in layers two through four and layers five through seven. The eighth layer is a one-dimensional convolutional layer with 1 filter and a kernel size of 3, employing the sigmoid activation function. The sigmoid layer ensures that all elements in the output vector are within the range [0,1], thus implementing a multi-label classification method.

[0078] First, in TensorFlow, we set up a bitstream s with randomly generated lengths K of 4, 32, and 64 bits, a noise z with randomly generated noise, and sample instantaneous channel state information (CSI) from the three channel sets, s B(1,0.5), and set the code rate. The training set A1 has a size of 1,000,000, and the test set A2 has a size of 10,000. Then, s obtains the transmitted signal x through the transmitter, and inputs x, random noise z, and instantaneous channel state information CSI into the AAR-GAN model to obtain a channel sample that approximates the real channel.

[0079] Next, in TensorFlow, we set up random generation of information bitstreams s with lengths K of 4, 32, and 64 bits, random generation of noise z, and sampling of instantaneous channel state information (CSI) from three channel sets, s B(1,0.5), and set the code rate. The training set B1 has a size of 1,000,000, and the test set B2 has a size of 10,000. Then, the transmitted signal x is obtained by the encoder of the R-CNN model, x is passed through the channel model to obtain the received signal y, and then y is passed through the decoder of the R-CNN model to obtain the recovered signal.

[0080] Then, in each iteration of training, the training set A1 is input into the AAR-GAN model, and the network is trained using the binary cross-entropy function and L2 regularization as loss functions, and the Adam optimization algorithm is used to train the AAR-GAN model. Then, the training set B1 is input into the RCNN model, and the network is trained using the binary cross-entropy function as loss function, and the Adam optimization algorithm is used to train the RCNN network model. The learning rate is set to α = 0.0001, the momentum is set to β1 = 0.5, and the number of training steps N is... D =N G =100, N T =N R =100, training batch size B=320. For AWGN channels, the number of iterative training rounds N iter =10, in E b Training an AAR-GAN model at / N0=12dB, and then... b Training the RCNN model at / N0=3dB; for the Rayleigh channel, instantaneous channel state information (CSI) needs to be sampled from the channel dataset, and the training iterations are N. iter =20, in E b Training an AAR-GAN model at / N0=38dB, and then training an E b The RCNN model is trained at / N0 = 20dB. For DeepMIMO channels, instantaneous channel state information (CSI) still needs to be sampled from the channel dataset. This channel dataset is obtained from the ray-traced DeepMIMO dataset. The training iterations are N. iter =20, in E b Training an AAR-GAN model at / N0=38dB, and then training an E bThe RCNN model was trained at / N0 = 20dB. All the models were trained independently and iteratively on their own.

[0081] When training the transmitter, the R-CNN model and the AAR-GAN residual generator are treated as a single network. The approximate gradients generated by the residual generator are used to train the transmitter, and finally, the end-to-end loss is calculated on the generated channel. When training the receiver, the R-CNN model and the real channel are treated as a single network. The real gradients of the real channel are used to train the receiver, and finally, the end-to-end loss is calculated on the real channel.

[0082] Testing begins after each training iteration, where the learned smart transceiver model is tested for BER performance on a real channel. For an AWGN channel, the number of training iterations is N. iter =10, in E b / N0 calculates the BER from 0dB to 7dB; for the Rayleigh channel, the number of iterative training rounds is N. iter =20, in E b / N0 calculates the BER from 0dB to 20dB; for DeepMIMO channels, the number of iterative training rounds is N. iter =20, in E b / N0 calculates BER from 0dB to 20dB.

[0083] like Figure 5 , Figure 6 and Figure 7 As shown, the simulation experiments of the above steps were conducted using PyCharm 2022 and TensorFlow 1.12 (Python 3.6) in the above embodiments, calculating and plotting different E values ​​under different channels. b BER comparison for / N0. In different E... b / N0, with E b With the increase of / N0, the intelligent transceiver design based on RCNN-GAN described in this invention (in) Figure 5 The text is a jumbled mix of seemingly unrelated phrases and sentences, making it impossible to translate coherently. It appears to be a collection of fragments from various sources, possibly related to RCNN-AAR-GAN, intelligent transceiver design, and possibly some other topics. Figure 5 The BER was calculated for AWGN, Rayleigh, and DeepMIMO channels (labeled as CNN-CGAN).

[0084] Several embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for designing an intelligent transceiver based on RCNN-GAN, characterized in that, Specifically, the steps include the following: S01, Steps for constructing a smart transceiver model; A residual-assisted generative adversarial network (AAR-GAN) model based on attention is used to simulate a channel. The AAR-GAN model includes a residual generator and a discriminator. The residual generator comprises three convolutional layers and a channel attention mechanism (SE-Block) network. The first layer consists of a convolutional layer, an SE-Block layer, a batch normalization (BN) layer, and a ReLU activation layer. The second layer consists of a convolutional layer, a BN layer, and a ReLU activation layer. The third layer consists of a convolutional layer. The SE-Block network includes compression and activation operation layers. The discriminator includes a network of two convolutional layers and ReLU activation layers and a network of one convolutional layer and sigmoid activation layer; The transmitter and receiver are built based on the residual autoencoder RCNN model; the RCNN model includes an encoder and a decoder, the encoder constitutes the transmitter and the decoder constitutes the receiver; The decoder comprises an eight-layer network. The first to seventh layers are all convolutional layers and ReLU activation layers. The eighth layer is a convolutional layer and a sigmoid activation layer. There are skip connections between the second convolutional layer and the fourth ReLU layer, and between the fifth convolutional layer and the seventh ReLU layer. S02, Steps for obtaining training and test sets; Randomly generate information bit stream s; randomly sample instantaneous channel state information CSI and noise z from the channel dataset; obtain the training set A1 and test set A2 of the AAR-GAN model and the training set B1 and test set B2 of the RCNN model; S03, Training Steps; The AAR-GAN model is trained using the training set A1, the parameters of the AAR-GAN model are updated, and the AAR-GAN model is iteratively trained until convergence. The RCNN model is trained using the training set B1, the parameters of the RCNN model are updated, and the RCNN model is trained iteratively until convergence. S04, Testing and Smart Transceiver Acquisition Steps; Input the test set A2 and the test set B2 into the converged AAR-GAN model and RCNN model to obtain the bit error rate (BER). If the BER is higher than the design standard value, continue iterating. If the BER is lower than the design standard value, end the iteration and obtain the intelligent transceiver.

2. The intelligent transceiver design method based on RCNN-GAN according to claim 1, characterized in that, In step S02, the channel dataset includes the AWGN channel dataset, the Rayleigh channel dataset, and the DeepMIMO channel dataset.

3. The intelligent transceiver design method based on RCNN-GAN according to claim 2, characterized in that, For the AWGN channel dataset, there is no need to sample instantaneous channel state information (CSI); for the Rayleigh channel dataset, the sampled instantaneous CSI follows a complex Gaussian distribution with a mean of 0 and a variance of 1; for the DeepMIMO channel dataset, the sampled instantaneous CSI is based on the ray tracing channel dataset.

4. The intelligent transceiver design method based on RCNN-GAN according to claim 3, characterized in that, In step S03, during the training process of the AAR-GAN model, the binary cross-entropy function and L2 regularization are used as loss functions.

5. The intelligent transceiver design method based on RCNN-GAN according to claim 4, characterized in that, In step S03, during the training of the RCNN model, the binary cross-entropy function is used as the loss function.

6. The intelligent transceiver design method based on RCNN-GAN according to claim 5, characterized in that, In step S03, the Adam algorithm is used to optimize the AAR-GAN model and the RCNN model.

Images