Communication system and method

Abstract

The present application provides a communication system and method, which relates to the field of communication technology. The method includes: inputting the first data to be sent into the first model of the transmitting end, and outputting a modulated signal corresponding to the data to be sent; combining the modulated signal output by the first model with the pilot signal into a transmission signal, and sending the transmission signal to the receiving end through a communication network; after the receiving end receives the received signal, restoring the received signal through the second model of the receiving end, and performing noise reduction processing on the restored signal through the third model of the receiving end to obtain the restored second data. The communication system and method provided by the present application are used to improve the ability of the receiving end of the wireless communication system to restore the original transmission signal sent by the transmitting end, while reducing the complexity of the model training process.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a communication system and method. Background Art

[0002] With the rapid development of mobile communications, more advanced signal detection technologies are needed to cope with a wider range of communication needs. In order to meet the communication needs between base stations and user equipment, wireless communication systems need to be able to accurately and effectively restore the original transmission signal sent by the transmitter at the receiving end in an end-to-end manner.

[0003] In related technologies, each function such as coding and modulation is implemented by a corresponding module, and the algorithm technology of each part is quite mature. However, better detection performance can only be achieved when the algorithm's prerequisites are highly matched with the usage scenario. The prior hypothesis model often deviates from the actual channel conditions, and for some scenarios with unknown channel conditions, traditional mathematical methods are difficult or even impossible to find the optimal solution.

[0004] Based on this, it is necessary to propose a more flexible and efficient global optimization method so that the wireless communication system can accurately and effectively restore the original transmission signal sent by the transmitter at the receiving end in an end-to-end manner. Summary of the invention

[0005] The purpose of this application is to provide a communication system and method for improving the ability of a receiving end of a wireless communication system to restore the original transmission signal sent by a sending end, while reducing the complexity of a model training process.

[0006] The present application provides a communication system, including: a transmitting end and a receiving end:

[0007] The transmitting end is used to input the first data to be sent into the first model, and output the modulation signal corresponding to the data to be sent; the transmitting end is also used to combine the modulation signal output by the first model with the pilot signal into a transmission signal, and send the transmission signal to the receiving end through the communication network; the receiving end is used to restore the received signal through the second model after receiving the received signal, and perform noise reduction processing on the restored signal through the third model to obtain the restored second data; wherein, the first model, the second model and the third model are obtained after training based on the two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the influence of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the influence of the channel on the transmission process and improve the signal restoration capability of the second model.

[0008] Optionally, the training process of the first training stage may include the following steps: inputting a first sample signal into a transmitting end model to obtain a modulated signal output by the transmitting end model; the first sample signal includes: a first preset number of transmitting sequences, each transmitting sequence having the same amount of data; superimposing additive noise on the frequency domain of the modulated signal output by the transmitting end model to obtain a first intermediate signal; inputting the first intermediate signal into a demodulation and decoding subnet model to obtain a first result signal; calculating the loss value of the binary cross entropy loss function of the first sample signal and the first result signal, and adjusting the model parameters of the transmitting end model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function; wherein the first model is a network model obtained after the transmitting end model is trained; and the third model is a network model obtained after the demodulation and decoding subnet model is trained.

[0009] Optionally, the training process of the second training stage may include the following steps: inputting a second sample signal into the first model to obtain a modulated signal output by the first model; the second sample signal includes: a second preset number of transmission sequences, each transmission sequence has the same amount of data; processing the modulated signal output by the transmitting end model according to the first preset processing method to obtain a real sample signal, and processing the modulated signal output by the transmitting end model according to the second preset processing method to obtain a third sample signal and a pilot signal; inputting the third sample signal as input information into the generator, and inputting the pilot signal as a conditional signal into the generator to output a first generated sample signal; inputting the real sample signal and the first generated sample signal into the discriminator, and based on the generator and the The comparative learning process of the discriminator adjusts the model parameters of the generator and the discriminator; wherein, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain a second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal; the second model is a network model obtained after the generator is trained.

[0010] Optionally, the step of inputting the real sample signal and the first generated sample signal into the discriminator and adjusting the model parameters of the generator and the discriminator based on the comparative learning process between the generator and the discriminator includes: inputting the real sample signal into the discriminator, calculating a first loss value of a binary cross entropy loss function of the real sample signal based on a sample label of the real sample signal, and adjusting the model parameters of the discriminator according to the first loss value; inputting the first generated sample signal into the discriminator, calculating a second loss value of a binary cross entropy loss function of the first generated sample signal based on a sample label of the first generated sample signal, and adjusting the model parameters of the discriminator according to the second loss value; generating a second generated sample signal based on a fourth sample signal, and inputting the second generated sample signal into the generator; calculating a third loss value of a binary cross entropy loss function of the second generated sample signal based on a sample label of the second generated sample signal, and adjusting the model parameters of the generator according to the third loss value; wherein the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

[0011] Optionally, the system further includes: deploying the first model to the sending end, and deploying the second model and the third model to the receiving end.

[0012] Optionally, the system also includes: after a channel of the communication network between the transmitting end and the receiving end changes, retraining the second model based on the changed channel and the training method corresponding to the second training stage; deploying the retrained second model to the receiving end to replace the old model on the receiving end.

[0013] The present application also provides a communication method, which is applied to a communication system, wherein the communication system includes: a transmitting end and a receiving end, and the method includes:

[0014] The first data to be sent is input into the first model of the transmitting end, and a modulation signal corresponding to the data to be sent is output; the modulation signal output by the first model is combined with the pilot signal into a transmission signal, and the transmission signal is sent to the receiving end through a communication network; after the receiving end receives the receiving signal, the receiving signal is restored by the second model of the receiving end, and the restored signal is subjected to noise reduction processing by the third model of the receiving end to obtain the restored second data; wherein, the first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the influence of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the influence of the channel on the transmission process and improve the signal restoration capability of the second model.

[0015] Optionally, before inputting the first data to be sent into the first model of the sending end and outputting the modulated signal corresponding to the data to be sent, the method also includes: inputting a first sample signal into the sending end model to obtain a modulated signal output by the sending end model; the first sample signal includes: a first preset number of sending sequences, each sending sequence has the same amount of data; additive noise is superimposed on the frequency domain of the modulated signal output by the sending end model to obtain a first intermediate signal; the first intermediate signal is input into the demodulation and decoding subnet model to obtain a first result signal; the loss value of the binary cross entropy loss function of the first sample signal and the first result signal is calculated, and the model parameters of the sending end model and the demodulation and decoding subnet model are adjusted according to the loss value of the binary cross entropy loss function; wherein the first model is a network model obtained after the sending end model is trained; and the third model is a network model obtained after the demodulation and decoding subnet model is trained.

[0016] Optionally, after adjusting the model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function, the method also includes: inputting a second sample signal into the first model to obtain a modulated signal output by the first model; the second sample signal includes: a second preset number of sending sequences, each sending sequence has the same amount of data; processing the modulated signal output by the transmitter model according to the first preset processing method to obtain a real sample signal, and processing the modulated signal output by the transmitter model according to the second preset processing method to obtain a third sample signal and a pilot signal; inputting the third sample signal as input information into the generator, and inputting the pilot signal as a conditional signal into the generator to output a first generated sample signal; combining the real sample signal and the first generated sample signal The signal is input into the discriminator, and the model parameters of the generator and the discriminator are adjusted based on the comparative learning process of the generator and the discriminator; wherein, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain a second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal; the second model is the network model obtained after the generator is trained.

[0017] Optionally, the step of inputting the real sample signal and the first generated sample signal into the discriminator and adjusting the model parameters of the generator and the discriminator based on the comparative learning process between the generator and the discriminator includes: inputting the real sample signal into the discriminator, calculating a first loss value of a binary cross entropy loss function of the real sample signal based on a sample label of the real sample signal, and adjusting the model parameters of the discriminator according to the first loss value; inputting the first generated sample signal into the discriminator, calculating a second loss value of a binary cross entropy loss function of the first generated sample signal based on a sample label of the first generated sample signal, and adjusting the model parameters of the discriminator according to the second loss value; generating a second generated sample signal based on a fourth sample signal, and inputting the second generated sample signal into the generator; calculating a third loss value of a binary cross entropy loss function of the second generated sample signal based on a sample label of the second generated sample signal, and adjusting the model parameters of the generator according to the third loss value; wherein the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

[0018] Optionally, after adjusting the model parameters of the generator according to the third loss value, the method further includes: deploying the first model to the sending end, and deploying the second model and the third model to the receiving end.

[0019] Optionally, after deploying the first model to the sending end, and deploying the second model and the third model to the receiving end, the method further includes: after a channel of the communication network between the sending end and the receiving end changes, retraining the second model based on the changed channel and the training method corresponding to the second training stage; deploying the retrained second model to the receiving end to replace the old model on the receiving end.

[0020] The present application also provides a computer program product, comprising a computer program / instruction, which implements the steps of any of the above-mentioned communication methods when executed by a processor.

[0021] The present application also 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 implements the steps of any of the above-mentioned communication methods when executing the program.

[0022] The present application also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned communication methods are implemented.

[0023] The communication system and method provided by the present application, after training the first model of the transmitting end, the second model and the third model of the receiving end based on the two-stage training method, can, when the transmitting end communicates with the receiving end, input the first data to be sent into the first model of the transmitting end, output the modulation signal corresponding to the data to be sent, and then combine the modulation signal output by the first model with the pilot signal into a transmission signal, and send the transmission signal to the receiving end through the communication network. Finally, after the receiving end receives the received signal, the received signal is restored by the second model of the receiving end, and the restored signal is subjected to noise reduction processing by the third model of the receiving end to obtain the restored second data. In this way, the wireless communication system can accurately and effectively restore the original transmission signal sent by the transmitting end at the receiving end. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 It is a structural diagram of a communication system in the related technology provided by this application;

[0026] Figure 2 It is a structural diagram of the communication system provided by this application;

[0027] Figure 3 It is a flowchart of the first training stage provided by this application;

[0028] Figure 4 is a flow chart of the second training phase provided by this application;

[0029] Figure 5 It is a flowchart of the communication method provided by this application;

[0030] Figure 6 It is a structural schematic diagram of the electronic device provided by this application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0032] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0033] In related technologies, the modular structures of the transmitter and receiver can be replaced by a complete neural network respectively. Commonly used networks are deep neural networks (DNN) or convolutional neural networks (CNN). The network at the transmitter is called an autoencoder, and the network at the receiver is called an autodecoder. Figure 1 As shown in the figure, the input of the sending network is the sending signal s, the output x is transmitted through the channel to obtain y, and the receiving network recovers the output according to y. It needs to be as close to s as possible to achieve a complete end-to-end communication network process.

[0034] For scenarios where the channel is known, existing studies have attempted to consider influencing factors such as additive white Gaussian noise (AWGN), carrier frequency offset, phase offset, sampling frequency, time offset, and multipath fading. However, such methods are still not practical enough for scenarios where the channel is unknown.

[0035] For scenarios with unknown channels, the end-to-end communication models based on neural networks have a common problem that needs to be solved: they cannot be reversely derived. That is, when training the transmitter network, because the back propagation needs to pass through an unknown channel, the necessary channel parameters are missing, and the network parameters cannot be updated by back propagation gradient derivation, resulting in the inability to complete the training of the transmitter network. For this problem, the existing solutions can be roughly divided into two categories: the first is to train a network as a differentiable proxy channel to approximate the distribution of the unknown channel; the second is to solve the approximate gradient through approximate calculation, and optimize the transmitter network without knowing the channel transfer function or CSI. In addition, there are some previous research works that only use neural networks to optimize some functions of both the transmitter and the receiver, such as channel estimation based on the concept of super-resolution reconstruction in the Orthogonal Frequency Division Multiplexing (OFDM) system, single-bit quantization channel estimation method based on DNN, modulation scheme based on constellation geometry and superimposed pilot, etc.

[0036] When the above technical solutions in the related art train the transmitter network under an unknown channel, because the back propagation needs to pass through an unknown channel, the necessary channel parameters are missing, and the network parameters cannot be updated by back propagation gradient derivation, resulting in the inability to reversely derive and the inability to complete the training of the transmitter network. The communication system and method provided in this application will be designed to solve the problem of being unable to derive in an unknown channel and to successfully train the transmitter network.

[0037] The above technical solutions in the related art only use the neural network structure in some modules, and fail to achieve complete end-to-end communication. Independent optimization limits the ability of the neural network to learn autonomously, and it is difficult to learn accurate features for complex and changeable unknown channel distributions. Therefore, the communication system and method provided by this application will be designed for the purpose of achieving a complete end-to-end neural network structure, learning accurate unknown channel features, and improving the global performance of the system.

[0038] The above technical solution in the related art is a model training that is iteratively completed by three modules: the transmitter, the unknown channel, and the receiver. After changing the channel scenario, the entire end-to-end model must be retrained, which takes a long time to retrain, has poor network flexibility, and no parameter update mechanism is designed. Therefore, the communication system and method provided by this application will be designed to reduce the number of parameters required for retraining, reduce time complexity and space complexity, and improve the flexibility and practicality of the solution.

[0039] The communication method provided in the embodiment of the present application is described in detail below through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0040] like Figure 2As shown, an embodiment of the present application provides a communication system, which includes: a transmitting end and a receiving end, wherein the transmitting end includes: a transmitting end network T (that is, the first model mentioned above); the receiving end includes: a generator G (that is, the second model mentioned above) and a demodulation and decoding subnet R2 (that is, the third model mentioned above).

[0041] Exemplarily, the transmitting end is used to input the first data to be sent into the first model, and output a modulated signal corresponding to the data to be sent; the transmitting end is also used to combine the modulated signal output by the first model and the pilot signal into a transmitting signal, and send the transmitting signal to the receiving end through a communication network; the receiving end is used to restore the receiving signal through the second model after receiving the receiving signal, and perform noise reduction processing on the restored signal through the third model to obtain the recovered second data.

[0042] Among them, the first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the impact of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the impact of the channel on the transmission process and improve the second model's ability to restore the signal.

[0043] For example, Figure 2 As shown, the transmitter network T and the receiver demodulation and decoding subnet R2 can be implemented using the CNN architecture, and the receiver generator and the discriminator D used to train the generator G can be implemented using the conditional generative adversarial network (CGAN) architecture. The sequence to be sent (i.e., the first data mentioned above) is denoted as s n , the modulation symbol output by the transmitter network T is recorded as X data . Let the pilot symbol be X p , the modulation symbol is combined with the pilot symbol to obtain the complete transmission signal, which is recorded as X. The received signal is recorded as Y, and the output of the generator G is recorded as The estimated transmission sequence recovered by the demodulation and decoding subnet R2 is recorded as

[0044] It is understandable that the transmitting antenna sends the signal in the form of electromagnetic waves through the air interface. During the transmission process, the electromagnetic waves are affected by the transmission environment, and the characteristics such as signal amplitude and phase will change randomly. Based on this, after the transmitted signal X passes through the real channel, the unknown fading of the channel will damage the signal, that is, the signal Y received by the receiving end is not exactly the same as the signal X sent by the transmitting end.

[0045] Exemplarily, the first training stage in the above two-stage training method is used to train the transmitting network T (i.e., the above first model) and the demodulation and decoding subnet R2 (i.e., the above third model) to improve the noise reduction capability of the demodulation and decoding subnet R2; the second training stage in the above two-stage training method is used to train the generator G (i.e., the above second model) to improve the signal restoration capability of the generator G.

[0046] Exemplarily, the training process of the first model, the second model, and the third model can be performed in stages on the same device. By dividing the model training process into two stages, when the channel of the communication network between the transmitter and the receiver changes, only the second training stage can be repeated to complete the update of the entire system to adapt to the new channel scenario. This greatly reduces the number of parameters required for retraining, reduces the time complexity and space complexity, and improves the flexibility and practicality of the solution.

[0047] Optionally, in an embodiment of the present application, the first training stage and the second training stage of the above two-stage training method are described in detail below.

[0048] Exemplarily, the training process of the first training stage may include the following steps 301 to 304:

[0049] Step 301: Input a first sample signal into a transmitter model to obtain a modulated signal output by the transmitter model.

[0050] The first sample signal includes: a first preset number of transmission sequences, and the data volume of each transmission sequence is the same.

[0051] Step 302: Additive noise is superimposed on the frequency domain of the modulated signal output by the transmitting end model to obtain a first intermediate signal.

[0052] Step 303: Input the first intermediate signal into a demodulation and decoding subnet model to obtain a first result signal.

[0053] Step 304: Calculate the loss value of the binary cross entropy loss function of the first sample signal and the first result signal, and adjust the model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function.

[0054] Among them, the first model is the network model obtained after the sending end model is trained; the third model is the network model obtained after the demodulation and decoding subnet model is trained.

[0055] For example, Figure 3FIG. 1 is a flow chart of the first training phase in the embodiment of the present application. In the first training phase, the transmitting end network T of the transmitting end and the demodulation and decoding subnet R2 of the receiving end are jointly trained. n After passing through the transmitting end network T, the modulated signal is output The dimensionality of the modulated signal matches the number of outputs at the transmitter when using traditional communication algorithms.

[0056] A noise layer is added between the transmitter network T and the demodulation and decoding subnetwork R2 to simulate the noise interference in actual communication. Through this architecture design, the transmitter network T and the receiver demodulation and decoding subnetwork R2 can have the ability to resist Gaussian noise. data After passing through the noise layer Send it to the demodulation and decoding subnet R2, and output an estimate of the transmitted sequence The network parameters are updated by calculating the binary cross entropy loss function so that the estimated sequence after multiple trainings is as close as possible to the original transmitted sequence. The loss function can be expressed as the following formula:

[0057]

[0058] Among them, s n Indicates the nth sequence sent, represents the nth sequence after the receiver recovers, and s n , Use the binary cross entropy loss function to calculate the original transmitted information bits s n (i.e. the first data above) and the data recovered by the receiving end (ie, the distance between the above-mentioned second data), and based on the loss value, update the model parameters of the sending end network T and the receiving end demodulation and decoding subnet R2.

[0059] Specifically, the first training stage may include the following steps A1 to A5:

[0060] Step A1: Initialize M transmission sequences with a length of k bits.

[0061] Step A2: During iteration, select the sequences s1 to s1 with the batch size batch_size in order. n , and are input into the transmitting network T in turn to output the modulated signal.

[0062] Step A3: Add additive noise data to the modulated signal, and then send it to the demodulation and decoding subnet R2 to output an estimate of the transmitted sequence.

[0063] Step A4: Calculate s n and The loss value of the binary cross entropy loss function is calculated, and based on the loss value, the optimization algorithm is used to update the model parameters of the sending network T and the demodulation and decoding subnetwork R2 to complete the training of a batch.

[0064] Step A5: Select the next batch_size sequence and repeat steps 2 to 4 for training. Repeat this process for iteration_size times to complete one epoch of training. A total of M sequences need to be trained for epoch_size times to achieve the desired T-R2 subnet performance.

[0065] Exemplarily, as shown in the following Table 1, the model parameters of each model in the first training stage are as follows:

[0066]

[0067] Table 1

[0068] Among them, 0.2 is the slope of the part of LeakRelu whose horizontal coordinate is less than 0, which is used to prevent the gradient from disappearing. The relevant configuration of the first training stage is shown in Table 2 below:

[0069] Loss Function BCE iteration 100 batch_size 1000 Optimization Algorithm Adam epoch 500 Learning Rate 0.01 Training SNR 5dB

[0070] Table 2

[0071] For example, the first training phase only considers Gaussian noise, but the real channel in the actual scenario often has more than just Gaussian noise, but also different types of fading, with unknown parameters that may change over time. Therefore, the second phase uses CGAN at the receiving end to specifically offset the impact of the channel on the transmitted signal, in order to achieve the function of estimation and equalization in the corresponding traditional communication.

[0072] Exemplarily, the training process of the second training stage may include the following steps 401 to 404:

[0073] Step 401: Input a second sample signal into the first model to obtain a modulated signal output by the first model.

[0074] The second sample signal includes: a second preset number of transmission sequences, each transmission sequence having the same amount of data.

[0075] Step 402: Process the modulated signal output by the transmitting end model according to a first preset processing method to obtain a real sample signal, and process the modulated signal output by the transmitting end model according to a second preset processing method to obtain a third sample signal and a pilot signal.

[0076] Step 403: Input the third sample signal as input information into a generator, and input the pilot signal as a conditional signal into the generator, and output a first generated sample signal.

[0077] Step 404: calculate a first loss value of the binary cross entropy loss function of the real sample signal and a second loss value of the binary cross entropy loss function of the first generated sample signal based on the sample label, and adjust the model parameters of the generator according to the first loss value, and adjust the model parameters of the discriminator according to the second loss value.

[0078] Among them, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain the second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal; the second model is a network model obtained after the generator is trained.

[0079] For example, in the second training phase, the estimated equalization subnetwork R1 (including the generator G and the discriminator D) at the receiving end is trained separately. First, the output X data As the initial input for this stage of training. Then, we need to prepare real samples that can be directly fed into the discriminator D, as well as generated samples that need to pass through the generator G and then be fed into the discriminator D. For real samples, X data X obtained by directly superimposing Gaussian noise N in the frequency domain data +N is the real sample. For the generated sample, let X data After the pilot signal passes through a channel sample H randomly selected from the fading channel database and then superimposed with the additive Gaussian noise data N in the frequency domain, the received signal Y can be obtained. The received signal Y is composed of the received data X data and receive pilot Y p The received data X in Y is data As input information, it is sent to the generator G, and the remaining received pilot Y p The conditional information is also fed into the generator G. The output of the generator G That is to generate samples.

[0080] Exemplarily, the real samples and generated samples are sent to the discriminator D, the discriminator outputs high / low values ​​corresponding to true / false, and the results of the discriminator D output are fed back to the generator G to update the model parameters of the generator G and the discriminator D. After multiple updates and iterations of adversarial learning of the generator G and the discriminator D, the generator G can finally confuse the discrimination ability of the discriminator D, so that the discriminator D cannot distinguish whether the sample comes from the output of the generator G or the real output, that is, the final sample discrimination probability is 0.5. At this time, the generator G in the CGAN can effectively resist the influence of the channel. When deployed, the received signal through the channel is input into the trained generator for channel estimation and equalization.

[0081] Specifically, the above step 404 may further include the following steps 404a1 to 404a4:

[0082] Step 404a1: input the real sample signal into the discriminator, calculate the first loss value of the binary cross entropy loss function of the real sample signal based on the sample label of the real sample signal, and adjust the model parameters of the discriminator according to the first loss value.

[0083] Step 404a2: input the first generated sample signal into the discriminator, calculate a second loss value of the binary cross entropy loss function of the first generated sample signal based on the sample label of the first generated sample signal, and adjust the model parameters of the discriminator according to the second loss value.

[0084] Step 404a3: Generate a second generated sample signal based on the fourth sample signal, and input the second generated sample signal into the generator.

[0085] Step 404a4: Calculate a third loss value of the binary cross entropy loss function of the second generated sample signal based on the sample label of the second generated sample signal, and adjust the model parameters of the generator according to the third loss value.

[0086] Among them, the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

[0087] It should be noted that the generation process of the second generated sample signal may refer to the generation process of the first generated sample signal.

[0088] Specifically, the second training stage may include the following steps B1 to A5:

[0089] Step B1: Initialize 2*M transmission sequences with a length of k bits.

[0090] Step B2: During iteration, select a sequence of D_batch_size in order, input them into the transmitting network T in sequence, and output the modulated signal.

[0091] Step B3: Add additive noise data to the modulated signal to obtain a real sample (ie, the real sample signal mentioned above).

[0092] Step B4: Send the real sample to the discriminator D, and set the label to true. Calculate the binary cross entropy loss function of the real sample and the true label, and use the optimization algorithm to update the network parameters of the discriminator.

[0093] Step B5: The modulated signal and pilot data outputted in step 2 are passed through a Rayleigh fading channel and additive noise data to obtain a received signal.

[0094] Step B6: Send the received signal to the generator G, output the generated sample (i.e., the first generated sample signal), and set the label to false. Calculate the binary cross entropy loss function of the generated sample and the false label, and use the optimization algorithm to update the network parameters of the discriminator. This completes the training of one batch of the discriminator D.

[0095] Step B7: Select the sequence of G_batch_size after the group D_batch_size to train the generator G. The modulated signal and pilot data output in step 3 are passed through the Rayleigh fading channel and additive noise data to obtain a received signal.

[0096] Step B8: Send the received signal to the generator G, output the generated sample, and set the label to true. Calculate the binary cross entropy loss function between the generated sample and the true label. Use the optimization algorithm to update the network parameters of the generator. This completes the training of one batch of the generator G.

[0097] Step B9, repeat steps 2 to 8, iterate iteration_size times, and complete one epoch of training. A total of 2*M sequences need to be trained epoch_size times to achieve the desired estimated balanced subnet R1 subnet performance. Using label smoothing during training can improve the discriminator's discrimination ability.

[0098] It should be noted that when inputting real samples to the discriminator D, X is selected data +N, not X data , for the following reasons:

[0099] First, the generator is used to remove the influence of the channel on the transmission signal, but it does not remove the noise. In order to make the real sample and the generated sample have more reasonable comparison conditions, the real sample sent to the generator should also contain noise. The work of eliminating the influence of noise is completed jointly by the sending end network and the demodulation and decoding subnet.

[0100] In addition, from the results of training T-R2 in stage 1, if there is no noise influence at all during training, even if the noise power is very small, it will be mistranslated during deployment, resulting in overfitting and poor model performance. Artificially adding some noise during training can avoid this situation.

[0101] Through the training of the above two training stages, the trained transmitter network T, channel estimation and equalization subnet R1 and demodulation and decoding subnet R2 are obtained. In the actual end-to-end communication process, the signal source is input to the transmitter network T to obtain the output, and the output of the transmitter network T is transmitted through the antenna as the transmission signal, and the signal is received by the receiver after passing through the real channel environment. The receiver first uses the generator G in the estimation and equalization subnet R1 to perform channel estimation and equalization on the received signal, and then inputs the result into the demodulation and decoding subnet R2 to demodulate and decode the equalized signal, and finally obtains the detection signal, thus completing a two-stage training transceiver end-to-end communication process based on a neural network.

[0102] Exemplarily, when the above sample label is true, the value of the sample label is a random number between 0.8 and 1. Using label smoothing can improve the discriminator's discriminative ability. The following Table 3 estimates the model parameters of each model in the balanced subnet R1:

[0103]

[0104] Table 3

[0105] Exemplarily, the training configuration of the second training stage is shown in Table 4 below, where the batch_size (i.e., D_batch_size) of the discriminator D is set slightly higher than the batch_size (i.e., G_batch_size) of the generator G, which helps to first improve the discrimination ability of the discriminator D, and then let the generator G fight against the discriminator D which already has a certain learning ability, which can make the samples generated by the generator G closer to the real samples. This prevents the discriminator D from being interfered by the generator G when it does not have sufficient discrimination ability, affecting the network performance. Setting this value is just a better solution found during the debugging process, and better configuration schemes are not excluded.

[0106]

[0107] Table 4

[0108] In a possible implementation, in order to be able to combine with OFDM technology in a multi-carrier scenario, based on Figure 2The structural design shown can also process the transmission signal X through the OFDM modulation module before transmitting the transmission signal. At the same time, at the receiving end, the received signal can also be processed through the OFDM demodulation module to improve the communication capability of the communication system in a multi-carrier scenario.

[0109] Optionally, in an embodiment of the present application, after completing the training of the above-mentioned model, the model needs to be deployed on corresponding devices to achieve communication between devices.

[0110] Exemplarily, after the above step 404, the communication system provided in the embodiment of the present application may further include the following step 405:

[0111] Step 405: deploy the first model to the sending end, and deploy the second model and the third model to the receiving end.

[0112] It can be understood that the transmitter needs to deploy the transmitter network T, that is, the first model mentioned above, and the receiver needs to deploy the generator and demodulation and decoding subnet R2, that is, the second model and the third model mentioned above, to achieve communication between the transmitter and the receiver.

[0113] It should be noted that the above transmitting end can also be used as a receiving end, and the above receiving end can also be used as a transmitting end. Based on this, in order to enable the device to have both transmitting and receiving capabilities, the above first model, second model and third model can be deployed on the same device at the same time.

[0114] Furthermore, in an embodiment of the present application, after the deployment of the model is completed, if the channel of the communication network between the transmitter and the receiver changes, the model needs to be retrained.

[0115] Exemplarily, after the above step 405, the communication system provided in the embodiment of the present application may further include the following steps 406 and 407:

[0116] Step 406: After a channel of the communication network between the transmitting end and the receiving end changes, the second model is retrained based on the changed channel and a training method corresponding to the second training stage.

[0117] Step 407: deploy the retrained second model to the receiving end to replace the old model on the receiving end.

[0118] Exemplarily, the fading channel database contains channel samples of multiple different channel distributions. (For example, one hundred groups of channel samples generated by Rayleigh distribution, one hundred groups of channel samples generated by Rice distribution) In the second training phase, a sample is randomly selected from the fading channel database for training. A sample can also be randomly selected for testing during subsequent testing. The channel samples used in each round of training can be different channel samples of the same channel part.

[0119] Based on this, when the channel distribution of the communication network between the transmitter and the receiver is different from the channel distribution of the channel samples used in the training process, the second model needs to be retrained based on the second training phase. In addition, since the channel samples are not used in the first training phase, the first model and the third model do not need to be retrained, which greatly reduces the time consumption of repeated training.

[0120] The communication system provided in the embodiment of the present application, after training the first model of the transmitting end, the second model and the third model of the receiving end based on the two-stage training method, can, when the transmitting end communicates with the receiving end, input the first data to be sent into the first model of the transmitting end, output the modulation signal corresponding to the data to be sent, and then combine the modulation signal output by the first model with the pilot signal into a transmission signal, and send the transmission signal to the receiving end through the communication network. Finally, after the receiving end receives the received signal, the received signal is restored by the second model of the receiving end, and the restored signal is subjected to noise reduction processing by the third model of the receiving end to obtain the restored second data. In this way, the wireless communication system can accurately and effectively restore the original transmission signal sent by the transmitting end at the receiving end.

[0121] It should be noted that the communication method provided in the embodiment of the present application can be executed by a communication device or a control module in the communication device for executing the communication method. In the embodiment of the present application, the communication device provided in the embodiment of the present application is described by taking the communication device executing the communication method as an example.

[0122] It should be noted that in the embodiments of the present application, the communication methods shown in the above-mentioned method drawings are all illustrated by taking one of the drawings in the embodiments of the present application as an example. In specific implementation, the communication methods shown in the above-mentioned method drawings can also be implemented in combination with any other drawings that can be combined as shown in the above-mentioned embodiments, which will not be repeated here.

[0123] The communication device provided in the present application is described below, and the communication method described below and the communication method described above can refer to each other.

[0124] Figure 5A flow chart of a communication method provided in an embodiment of the present application is applied to a communication system, wherein the communication system includes: a transmitting end and a receiving end, such as Figure 5 As shown, the method specifically includes the following steps 501 to 503:

[0125] Step 501: input first data to be sent into the first model of the transmitting end, and output a modulated signal corresponding to the data to be sent.

[0126] Step 502: Combine the modulated signal output by the first model and the pilot signal into a transmission signal, and send the transmission signal to the receiving end through a communication network.

[0127] Step 503: After the receiving end receives the received signal, the received signal is restored through the second model of the receiving end, and the restored signal is subjected to noise reduction processing through the third model of the receiving end to obtain restored second data.

[0128] Among them, the first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the impact of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the impact of the channel on the transmission process and improve the second model's ability to restore the signal.

[0129] Optionally, before inputting the first data to be sent into the first model of the sending end and outputting the modulated signal corresponding to the data to be sent, the method also includes: inputting a first sample signal into the sending end model to obtain a modulated signal output by the sending end model; the first sample signal includes: a first preset number of sending sequences, each sending sequence has the same amount of data; additive noise is superimposed on the frequency domain of the modulated signal output by the sending end model to obtain a first intermediate signal; the first intermediate signal is input into the demodulation and decoding subnet model to obtain a first result signal; the loss value of the binary cross entropy loss function of the first sample signal and the first result signal is calculated, and the model parameters of the sending end model and the demodulation and decoding subnet model are adjusted according to the loss value of the binary cross entropy loss function; wherein the first model is a network model obtained after the sending end model is trained; and the third model is a network model obtained after the demodulation and decoding subnet model is trained.

[0130] Optionally, after adjusting the model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function, the method also includes: inputting a second sample signal into the first model to obtain a modulated signal output by the first model; the second sample signal includes: a second preset number of sending sequences, each sending sequence has the same amount of data; processing the modulated signal output by the transmitter model according to the first preset processing method to obtain a real sample signal, and processing the modulated signal output by the transmitter model according to the second preset processing method to obtain a third sample signal and a pilot signal; inputting the third sample signal as input information into the generator, and inputting the pilot signal as a conditional signal into the generator to output a first generated sample signal; combining the real sample signal and the first generated sample signal The signal is input into the discriminator, and the model parameters of the generator and the discriminator are adjusted based on the comparative learning process of the generator and the discriminator; wherein, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain a second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal; the second model is the network model obtained after the generator is trained.

[0131] Optionally, the step of inputting the real sample signal and the first generated sample signal into the discriminator and adjusting the model parameters of the generator and the discriminator based on the comparative learning process between the generator and the discriminator includes: inputting the real sample signal into the discriminator, calculating a first loss value of a binary cross entropy loss function of the real sample signal based on a sample label of the real sample signal, and adjusting the model parameters of the discriminator according to the first loss value; inputting the first generated sample signal into the discriminator, calculating a second loss value of a binary cross entropy loss function of the first generated sample signal based on a sample label of the first generated sample signal, and adjusting the model parameters of the discriminator according to the second loss value; generating a second generated sample signal based on a fourth sample signal, and inputting the second generated sample signal into the generator; calculating a third loss value of a binary cross entropy loss function of the second generated sample signal based on a sample label of the second generated sample signal, and adjusting the model parameters of the generator according to the third loss value; wherein the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

[0132] Optionally, after adjusting the model parameters of the generator according to the third loss value, the method further includes: deploying the first model to the sending end, and deploying the second model and the third model to the receiving end.

[0133] Optionally, after deploying the first model to the sending end, and deploying the second model and the third model to the receiving end, the method further includes: after a channel of the communication network between the sending end and the receiving end changes, retraining the second model based on the changed channel and the training method corresponding to the second training stage; deploying the retrained second model to the receiving end to replace the old model on the receiving end.

[0134] The communication method provided by the present application, after training the first model of the transmitting end, the second model and the third model of the receiving end based on the two-stage training method, can, when the transmitting end communicates with the receiving end, input the first data to be sent into the first model of the transmitting end, output the modulation signal corresponding to the data to be sent, and then combine the modulation signal output by the first model with the pilot signal into a transmission signal, and send the transmission signal to the receiving end through the communication network. Finally, after the receiving end receives the received signal, the received signal is restored by the second model of the receiving end, and the restored signal is subjected to noise reduction processing by the third model of the receiving end to obtain the restored second data. In this way, the wireless communication system can accurately and effectively restore the original transmission signal sent by the transmitting end at the receiving end.

[0135] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6 As shown, the electronic device may include: a processor (processor) 610, a communication interface (Communications Interface) 620, a memory (memory) 630 and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call the logic instructions in the memory 630 to execute the communication method, which includes: inputting the first data to be sent into the first model of the transmitting end, and outputting the modulation signal corresponding to the data to be sent; merging the modulation signal output by the first model and the pilot signal into a transmission signal, and sending the transmission signal to the receiving end through the communication network; after the receiving end receives the received signal, the received signal is restored by the second model of the receiving end, and the restored signal is subjected to noise reduction processing by the third model of the receiving end to obtain the restored second data; wherein the first model, the second model and the third model are obtained after training based on the two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capability of the third model and reduce the influence of noise on the demodulation and decoding capability of the third model; the second training stage is used to train the second model to reduce the influence of the channel on the transmission process and improve the signal restoration capability of the second model.

[0136] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0137] On the other hand, the present application also provides a computer program product, which includes a computer program stored on a computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the communication method provided by the above methods, and the method includes: inputting the first data to be sent into the first model of the sending end, and outputting a modulated signal corresponding to the data to be sent; combining the modulated signal output by the first model with the pilot signal into a sending signal, and sending the sending signal to the receiving end through a communication network; after the receiving end receives the receiving signal, the receiving end is modulated by the second model of the receiving end. The signal is restored, and the restored signal is subjected to noise reduction processing through the third model of the receiving end to obtain the recovered second data; wherein, the first model, the second model and the third model are obtained after training based on the two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the influence of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the influence of the channel on the transmission process and improve the signal restoration capability of the second model.

[0138] On the other hand, the present application also provides a computer-readable storage medium having a computer program stored thereon, which is implemented by a processor to execute the above-mentioned communication methods, the method comprising: inputting the first data to be sent into the first model of the transmitting end, and outputting a modulated signal corresponding to the data to be sent; merging the modulated signal output by the first model with the pilot signal into a transmitting signal, and sending the transmitting signal to the receiving end through a communication network; after the receiving end receives the receiving signal, restoring the receiving signal through the second model of the receiving end, and performing noise reduction processing on the restored signal through the third model of the receiving end to obtain the restored second data; wherein the first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method comprises: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capability of the third model and reduce the influence of noise on the demodulation and decoding capability of the third model; the second training stage is used to train the second model to reduce the influence of the channel on the transmission process and improve the signal restoration capability of the second model.

[0139] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0140] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A communication system, characterized in that: Includes: sending end and receiving end: The transmitting end is used to input first data to be sent into a first model, and output a modulated signal corresponding to the data to be sent; The transmitting end is further used to combine the modulated signal output by the first model and the pilot signal into a transmitting signal, and send the transmitting signal to the receiving end through a communication network; The receiving end is used to restore the received signal through the second model after receiving the received signal, and perform noise reduction processing on the restored signal through the third model to obtain restored second data; The first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the impact of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the impact of the channel on the transmission process and improve the second model's ability to restore the signal; The training process of the first training stage may include the following steps: Inputting a first sample signal into a transmitting end model to obtain a modulated signal output by the transmitting end model; the first sample signal includes: a first preset number of transmitting sequences, each transmitting sequence having the same amount of data; Additive noise is superimposed on the frequency domain of the modulated signal output by the transmitting end model to obtain a first intermediate signal; Inputting the first intermediate signal into a demodulation and decoding subnet model to obtain a first result signal; Calculating a loss value of a binary cross entropy loss function of the first sample signal and the first result signal, and adjusting model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function; Among them, the first model is the network model obtained after the transmitter model is trained; the third model is the network model obtained after the demodulation and decoding subnet model is trained; and the second model is the network model obtained after the generator is trained.

2. The system according to claim 1, characterized in that The training process of the second training stage may include the following steps: Inputting a second sample signal into the first model to obtain a modulated signal output by the first model; the second sample signal includes: a second preset number of transmission sequences, each transmission sequence having the same amount of data; Processing the modulated signal output by the transmitting end model according to a first preset processing method to obtain a real sample signal, and processing the modulated signal output by the transmitting end model according to a second preset processing method to obtain a third sample signal and a pilot signal; Inputting the third sample signal as input information into a generator, and inputting the pilot signal as a conditional signal into the generator, and outputting a first generated sample signal; Inputting the real sample signal and the first generated sample signal into a discriminator, and adjusting model parameters of the generator and the discriminator based on a comparative learning process between the generator and the discriminator; Among them, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain the second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal.

3. The system according to claim 2, characterized in that The step of inputting the real sample signal and the first generated sample signal into a discriminator, and adjusting the model parameters of the generator and the discriminator based on a comparative learning process between the generator and the discriminator, comprises: Inputting the real sample signal into the discriminator, calculating a first loss value of a binary cross entropy loss function of the real sample signal based on a sample label of the real sample signal, and adjusting a model parameter of the discriminator according to the first loss value; Inputting the first generated sample signal into the discriminator, calculating a second loss value of a binary cross entropy loss function of the first generated sample signal based on a sample label of the first generated sample signal, and adjusting a model parameter of the discriminator according to the second loss value; generating a second generated sample signal based on the fourth sample signal, and inputting the second generated sample signal into the generator; Calculating a third loss value of the binary cross entropy loss function of the second generated sample signal based on the sample label of the second generated sample signal, and adjusting the model parameters of the generator according to the third loss value; Among them, the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

4. The system according to claim 2 or 3, characterized in that: The system further comprises: The first model is deployed to the sending end, and the second model and the third model are deployed to the receiving end.

5. The system according to claim 4, characterized in that The system further comprises: After a channel of the communication network between the transmitting end and the receiving end changes, retraining the second model based on the changed channel and a training method corresponding to the second training stage; The retrained second model is deployed to the receiving end to replace the old model on the receiving end.

6. A communication method, characterized in that: Applied to a communication system, the communication system comprises: a transmitting end and a receiving end, the method comprises: Inputting first data to be sent into the first model of the transmitting end, and outputting a modulated signal corresponding to the data to be sent; Combining the modulated signal output by the first model with the pilot signal into a transmission signal, and sending the transmission signal to the receiving end through a communication network; After the receiving end receives the received signal, the receiving end restores the received signal through the second model of the receiving end, and performs noise reduction processing on the restored signal through the third model of the receiving end to obtain restored second data; The first model, the second model and the third model are obtained after training based on a two-stage training method; the two-stage training method includes: a first training stage and a second training stage; the first training stage is used to train the first model and the third model to improve the demodulation and decoding capabilities of the third model and reduce the impact of noise on the demodulation and decoding capabilities of the third model; the second training stage is used to train the second model to reduce the impact of the channel on the transmission process and improve the second model's ability to restore the signal; Before inputting the first data to be sent into the first model of the transmitting end and outputting a modulated signal corresponding to the data to be sent, the method further includes: Inputting a first sample signal into a transmitting end model to obtain a modulated signal output by the transmitting end model; the first sample signal includes: a first preset number of transmitting sequences, each transmitting sequence having the same amount of data; Additive noise is superimposed on the frequency domain of the modulated signal output by the transmitting end model to obtain a first intermediate signal; Inputting the first intermediate signal into a demodulation and decoding subnet model to obtain a first result signal; Calculating a loss value of a binary cross entropy loss function of the first sample signal and the first result signal, and adjusting model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function; Among them, the first model is the network model obtained after the transmitter model is trained; the third model is the network model obtained after the demodulation and decoding subnet model is trained; and the second model is the network model obtained after the generator is trained.

7. The method according to claim 6, characterized in that After adjusting the model parameters of the transmitter model and the demodulation and decoding subnet model according to the loss value of the binary cross entropy loss function, the method further includes: Inputting a second sample signal into the first model to obtain a modulated signal output by the first model; the second sample signal includes: a second preset number of transmission sequences, each transmission sequence having the same amount of data; Processing the modulated signal output by the transmitting end model according to a first preset processing method to obtain a real sample signal, and processing the modulated signal output by the transmitting end model according to a second preset processing method to obtain a third sample signal and a pilot signal; Inputting the third sample signal as input information into a generator, and inputting the pilot signal as a conditional signal into the generator, and outputting a first generated sample signal; Inputting the real sample signal and the first generated sample signal into a discriminator, and adjusting model parameters of the generator and the discriminator based on a comparative learning process between the generator and the discriminator; Among them, the first preset processing method includes: mixing additive noise into the modulated signal output by the transmitting end model to obtain a second intermediate signal; the second preset processing method includes: merging the modulated signal output by the transmitting end model with the pilot signal, and passing the combined signal through any sample channel in the fading channel database to obtain the second intermediate signal; the second preset processing method also includes: superimposing additive noise on the frequency domain of the second intermediate signal to obtain a third intermediate signal, and using the received data in the third intermediate signal as the third sample signal; the third intermediate signal includes: the received data and the pilot signal.

8. The method according to claim 7, characterized in that The step of inputting the real sample signal and the first generated sample signal into a discriminator, and adjusting the model parameters of the generator and the discriminator based on a comparative learning process between the generator and the discriminator, comprises: Inputting the real sample signal into the discriminator, calculating a first loss value of a binary cross entropy loss function of the real sample signal based on a sample label of the real sample signal, and adjusting a model parameter of the discriminator according to the first loss value; Inputting the first generated sample signal into the discriminator, calculating a second loss value of a binary cross entropy loss function of the first generated sample signal based on a sample label of the first generated sample signal, and adjusting a model parameter of the discriminator according to the second loss value; generating a second generated sample signal based on the fourth sample signal, and inputting the second generated sample signal into the generator; Calculating a third loss value of the binary cross entropy loss function of the second generated sample signal based on the sample label of the second generated sample signal, and adjusting the model parameters of the generator according to the third loss value; Among them, the sample label of the real sample signal is true; the sample label of the first generated sample signal is false; and the sample label of the second generated sample signal is true.

Images