End-to-end optimization method, constellation geometry and probability joint shaping method and communication device

Abstract

The present application provides an end-to-end optimization method, a constellation geometry and probability joint shaping method, and a communication device. In the end-to-end optimization method, a neural network is used to construct a transmitter of a communication system. The transmitted signal encoded by the transmitter is transmitted in a real channel and then reaches the receiving end. The received signal first passes through a trainable equalizer connected to the front end of the receiver to perform channel estimation and equalization on the received signal. The equalized signal is affected by quasi-additive noise and is input into the receiver for demodulation, decoding, and recovery of the original information. The receiver is implemented using a neural network or a digital signal processing algorithm. The end-to-end optimization framework of the autoencoder neural network includes the transmitter, the real channel, the trainable equalizer, and the receiver. The equalizer, transmitter, and receiver in the end-to-end optimization framework are trained. This method can implement the geometric and probabilistic joint shaping of the signal constellation in the communication system.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent communication technology, and in particular relates to an end-to-end optimization method, a constellation geometry and probability joint shaping method, and a communication device. Background Art

[0002] Artificial Intelligence (AI)-based methods have significantly enhanced the performance of traditional communication systems by independently optimizing coding, detection, channel estimation, equalization, and other communication components. Since independent optimization of each communication component cannot guarantee global optimal performance, end-to-end (E2E) optimization of communication systems based on deep learning (DL) within AI has pioneered a new paradigm for global optimization of communication systems. The transmitter and receiver are constructed using a neural network (NN), and the communication channel is modeled as a gradient-transferable model. This NN transmitter (T-NN), communication channel model, and receiver NN (R-NN) are sequentially concatenated and encapsulated into an autoencoder (AE) neural network, enabling adaptive E2E optimization for specific performance metrics and channel models within a unified neural network framework. This AE-based E2E optimization technology has been validated theoretically and experimentally in various communication systems, including those based on wireless, fiber, and fiber-wireless converged channels.

[0003] In related technologies, the training of AE neural networks uses the backpropagation (BP) algorithm. T-NN and R-NN optimize signal encoding and decoding through iterative training to combat the damage caused by the channel model, thereby improving the performance of the communication system after E2E optimization.

[0004] In related technologies, the key to enabling end-to-end optimization of communication systems using AE neural networks lies in the fact that the channel model connecting the T-NN and R-NN must provide a path for gradient propagation, meaning the channel model must be differentiable. Furthermore, this differentiable channel model must accurately simulate various impairments in real channels, such as multipath, nonlinear effects, and bandwidth limitations. This is necessary to ensure that the optimized encoding and decoding methods of the T-NN and R-NN improve the performance of end-to-end communication systems in real channels.

[0005] In related technologies, to accurately simulate real-world channel impairments, differentiable channel models within AE neural networks often rely on complex modeling algorithms or require the collection of extensive channel transmit and receive data to train a neural network channel model that can simulate channel responses. Regardless of the modeling approach, training a channel model that can be embedded within the AE before training the neural network requires significant computational resources while ensuring sufficient accuracy.

[0006] In related technologies, since the channel model cannot accurately reflect the various impairments of the real channel, the joint signal optimization method of using AE neural network to perform geometric constellation shaping (GS) and probabilistic constellation shaping (PS) on the constellation of communication signals is still in the simulation stage and difficult to be deployed in a real system. Summary of the Invention

[0007] To overcome the above problems, the present application provides an end-to-end optimization method, a constellation geometry and probability joint shaping method, and a communication device, which specifically include:

[0008] In a first aspect, the present application provides a method for end-to-end intelligent optimization of a communication system without channel model assistance, which is performed by a system-level end-to-end intelligent optimization framework. The method includes:

[0009] A communication system's transmitter, a T-NN, is constructed using a neural network. The T-NN-encoded transmitted signal travels through a real channel before reaching the receiver. The received signal first passes through a trainable equalizer connected to the receiver's front end to perform channel estimation and equalization. When the channel estimation and equalization are sufficiently accurate, the equalized signal appears to be affected only by additive noise. This noise-influenced signal is then demodulated and decoded in the receiver to recover the original information. The receiver can be implemented using a R-NN neural network or a digital signal processing (DSP) algorithm. The T-NN, the real channel, the trainable equalizer, and the receiver form an E2E (E2E) optimization framework using an AE neural network. In this framework, the error in the equalizer's channel estimation and equalization and the error after decoding by the receiver constitute the total loss function of the E2E optimization framework. A backpropagation algorithm updates the parameters of the trainable components of the E2E optimization framework based on the gradient of the total loss function. This results in improved encoding performance in the updated T-NN, better channel estimation and equalization performance in the updated equalizer, and improved decoding of the equalized signal by the updated R-NN when the receiver uses an R-NN.

[0010] The solution presented in this application enables an equalizer to accurately estimate the channel of a received signal transmitted in a real channel, leaving only the residual effects of additive noise in the equalized signal. T-NN, a trainable equalizer, and a receiver can use a unified loss function within an end-to-end optimization framework to perform gradient propagation and parameter updates in the real channel, eliminating the need to construct a complex microchannel model in advance and conserving computational resources.

[0011] In one possible implementation, the transmission data generated by the T-NN is divided into leading training data and payload. The leading training data participates in the calculation of the total loss of the AE neural network and updates the parameters of the trainable portion of the AE neural network via the BP algorithm. During the initialization phase of the end-to-end communication system, the payload is used to calculate the receiver decoding error rate to characterize the training effect and trigger the training termination condition. During the deployment and operation phase of the end-to-end communication system, it directly carries communication data for end-to-end communication. The proportion of training data and payload in the transmission data is dynamically adjusted based on the performance of the end-to-end communication system on the payload. When the payload can be recovered at a low bit error rate at the receiver, the proportion of training data in the transmission data decreases. When the payload decoding error at the receiver is large, the proportion of training data in the transmission data increases, thereby strengthening the training of the AE neural network and reducing the decoding error of future received data.

[0012] The scheme shown in this application, the transmission data structure of the leading training data and the payload enables the end-to-end communication system to perform communication and end-to-end optimization processes simultaneously online, ensuring the stability of the end-to-end communication system and its responsiveness to channel mutations.

[0013] In a second aspect, the present application provides a method for optimizing signal constellation geometry and probability in an end-to-end intelligent communication system, which is a method for optimizing communication signals. The method includes:

[0014] The end-to-end intelligent communication system that implements the joint shaping of constellation geometry and probability shown in the present application uses the intelligent optimization method without channel model assistance proposed in the first aspect for training. The end-to-end intelligent communication system that implements the joint shaping of signal constellation geometry and probability shown in the present application is also an AE neural network composed of T-NN, a real channel, a trainable equalizer and a receiver. T-NN is designed in a modular way, and each module includes a constellation probability generator, a sampler, a one-hot code to bit vector mapper (One-hot to Bit-vector Mapper, OBM), a bit vector to constellation geometry position mapper (Bit-vector to Constellation Mapper, BCM) and a modulator. The constellation probability generator outputs a probability vector, which contains the probability configuration of each constellation point. The sampler generates one-hot coded symbols based on probability vector sampling. These symbols correspond one-to-one to constellation points in the constellation diagram, achieving probabilistic constellation shaping (PS). These symbols are converted into bit vectors by the OBM. Different bit vectors are input into the BCM neural network, where they are mapped to constellation points at different locations, achieving geometric constellation shaping (GS). The symbols after PS and GS are fed into the modulator to generate the final transmit waveform (Tx). The transmit waveform is transmitted over a real channel. At the receiver, a trainable equalizer is used for channel estimation and equalization. After equalization, the signal is only affected by residual additive noise. The receiver can use a R-NN neural network or a digital signal processing (DSP) algorithm. After demodulation and decoding at the receiver, the error between the recovered bit information and the original transmitted bit information can be calculated. The recovery error and the equalization error of the equalizer together constitute the total loss function of the AE neural network. Based on this loss function, the BP algorithm updates the parameters of the trainable probability generator, BCM, and equalizer in the end-to-end intelligent communication system, thereby realizing the joint optimization of the transceiver and transmitter of the communication system.

[0015] The scheme presented in this application utilizes both probabilistic shaping and geometric shaping in a modularized manner within the T-NN transmitter, and is simultaneously optimized using the BP algorithm. This greatly exploits the freedom of shaping the signal constellation, effectively controlling the position and probability of occurrence of different constellation points to better reduce the impact of channel distortion on signal recovery and enhance the communication system receiver's ability to correctly decode signals. Geometric shaping is implemented using a neural network whose input is a bit vector and whose output is the position of the constellation point. While adjusting the geometric position of the constellation point, the bit labels corresponding to the constellation point are also optimized, ensuring that when demapping the constellation of the received signal, misjudgment of the constellation point results in the fewest bit errors, thereby optimizing the bit decoding performance of the optimized end-to-end intelligent communication system.

[0016] In one possible implementation, to ensure gradient propagation throughout the entire T-NN process, from the constellation probability generator generating the PS probability vector to the final Tx waveform generation, the sampler uses the Gumbel-Softmax method to generate one-hot encoded symbols consistent with the probability vector through a Gumbel sampler, ensuring the gradient of the total loss function with respect to the probability vector. The one-hot encoded symbols are converted into bit vectors by a pre-trained and fixed neural network (OBM). The OBM implemented via the neural network also ensures that it can be easily embedded in the T-NN for gradient propagation. To ensure that the neural network (BCM) that maps bit vectors to constellation points can stably and quickly reach optimality during end-to-end optimization training without the assistance of a channel model, the BCM is pre-trained using the standard Gray mapping rule for bit-to-lattice constellations before being embedded in the T-NN. This pre-training scheme provides the BCM with a suboptimal starting point, allowing it to converge to the optimal bit-to-constellation (GS) configuration more quickly than with random parameter initialization.

[0017] According to the solution shown in this application, pre-training only needs to be performed once during the training initialization phase of the AE neural network, which can significantly accelerate the training convergence speed of the AE neural network and reduce the computing resource consumption generated during the training process.

[0018] In a third aspect, the present application provides a neural network-based communication device, which has the function of implementing the first aspect or any optional method of the first aspect. The device includes at least one module, and the at least one module is used to implement the channel model-free end-to-end communication system intelligent optimization method provided by the first aspect or any optional method of the first aspect.

[0019] In a fourth aspect, the present application provides a neural network-based communication device, which has the functionality to implement the second aspect or any optional embodiment of the second aspect. The device includes at least one module, which is configured to implement the signal constellation geometry and probability joint shaping optimization method in a communication system provided by the second aspect or any optional embodiment of the second aspect.

[0020] Using the solution of the present invention, end-to-end optimized communication systems based on neural networks no longer require pre-training or building a channel model that accurately estimates all channel characteristics. This method can implement a joint training process for the neural network transceiver in an end-to-end intelligent communication system and achieve joint geometric and probabilistic shaping of the signal constellation in the communication system. This technology reduces the computational cost of training the channel model in an end-to-end intelligent communication system, improves the system's adaptability to channel dynamics, and reduces the actual deployment cost of the end-to-end intelligent communication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a schematic diagram of an end-to-end intelligent communication system architecture based on a constellation geometry and probability joint shaping optimization technology without channel model assistance provided by an exemplary embodiment of the present application.

[0022] Figure 2 It is a flowchart of an end-to-end intelligent optimization method for a communication system without channel model assistance provided by an exemplary embodiment of the present application.

[0023] Figure 3 It is a schematic diagram of the architecture of a transmitter provided by an exemplary embodiment of the present application.

[0024] Figure 4 It is a schematic diagram of the architecture of a receiver provided by an exemplary embodiment of the present application.

[0025] Figure 5 This is the performance result of an exemplary embodiment of the present application provided under different received optical powers in an actual optical fiber communication system. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0027] The following is an explanation of some terminology concepts involved in the embodiments of this application.

[0028] 1. AE Neural Network: A neural network is a computational model composed of a large number of nodes (or "neurons") connected by weighted connections. It can be trained to learn and simulate complex nonlinear relationships. An autoencoder is an unsupervised neural network that learns a compressed representation of data through input-to-output transformations. In communication systems, AE neural networks use self-learning mechanisms to adaptively adjust signal transmission processes, optimize signal quality, and thus reduce the impact of channel losses on the system information rate.

[0029] 2. E2E Optimization and E2E Communication Systems: E2E optimization refers to optimizing the transmitter and receiver as a combined system within an overall communication system, often with the goal of maximizing overall performance, such as the transmission rate or reliability of the communication system. An E2E communication system considers and designs the entire transmission link as a single system, rather than dividing it into multiple, independently optimized components. Neural networks can be used to optimize signal processing and improve overall communication quality.

[0030] 3. GS: In communication systems, a constellation diagram is a graphic used to represent signals. GS refers to adjusting the positions of constellation points to optimize system performance, such as improving noise tolerance by increasing the minimum distance between constellation points.

[0031] 4. PS: Constellation PS involves modifying the probability of occurrence of different constellation points in the transmitted signal to further optimize the information rate and improve the overall performance of the communication system. For example, by optimizing the probability distribution to approach the channel capacity, the transmission rate can be increased. In linear additive white Gaussian noise (AWGN) channels, the Maxwell-Boltzmann (MB) distribution can be applied to achieve an information rate close to the theoretical optimal PS. However, for channels with nonlinearity and memory, the MB-distributed PS signal cannot guarantee the optimal information rate. Optimal PS strategies in nonlinear and memory-prone channels require further research and exploration.

[0032] The following describes the relevant background of the embodiments of the present application.

[0033] In communication systems, the concept of end-to-end (E2E) optimization has pioneered a new paradigm in which the transmitter and receiver are modeled as neural networks (NNs), while the communication channel is treated as an untrainable intermediate layer. This configuration is then encapsulated in an autoencoder (AE), allowing targeted optimization for specific performance metrics and channel models within a unified framework. Numerous simulation studies have demonstrated that joint optimization of signal quality (GS) and signal quality (PS) using E2E optimization techniques improves communication rate and signal impairment mitigation compared to conventional MB-distributed probabilistically shaped signals in both wireless and wired communication systems. However, these joint optimization methods remain at the simulation stage, with no real-world experimental cases demonstrating their feasibility. While the theoretical framework of E2E optimization is attractive, it faces significant practical obstacles, particularly the requirement for a known and differentiable channel model. This presents a significant challenge, as not all channel models support differentiation—for example, quantization within the transceiver is a prime example of this limitation, which disrupts gradient-based training via backpropagation. Furthermore, research has shown that reliance on simplified differentiable models can introduce performance inaccuracies when transitioning from numerical simulations to real channels. To achieve this precision, accurate differentiable channel models can become prohibitively complex, imposing a significant computational burden and complicating the optimization process. To address non-differentiable or non-analytic channel characteristics, generative adversarial networks (GANs) have emerged as a potential channel model construction method and have been applied to end-to-end (E2E) optimization in communication systems. However, introducing GANs into the channel model adds an additional stage to the AE neural network optimization process, requiring iterative acquisition of channel samples to update the GAN—a data-intensive and time-consuming process. Furthermore, given that the data generated by GANs primarily describes the deterministic characteristics of the real channel (such as the channel impulse response), they cannot provide accurate noise estimates. Therefore, AE neural networks constructed using GANs struggle to ensure the accuracy and reliability of PS optimization. Recent work has demonstrated the potential of gradient-free methods, but these gradient-free optimization approaches separate the AE neural network into independent transmitter and receiver neural networks, optimizing each separately. This strategy violates the fundamental principle of E2E optimization—exploiting the benefits of joint optimization of the system's transmitter and receiver, as it fails to optimize both the transmitter and receiver simultaneously and requires more samples and iterations to achieve convergence. The training cost and instability result in low reliability and high deployment costs for this method.

[0034] In this application, an additional equalization loss is introduced into the training loss function in the AE neural network, and a trainable equalizer is accurately optimized to estimate and compensate for deterministic channel distortion, thereby limiting the residual signal damage that the subsequent decoder needs to process to quasi-additive noise. Therefore, the decoder interprets the equalized signal as being affected only by the additive noise, enabling it to perform bit metric decoding efficiently and simply. During the E2E optimization process of the neural transceiver, the decoding error of the receiver constitutes a part of the loss of the training process, guiding the joint learning of PS and GS. This progress enables the AE neural network (GPS-AE) with the ability to jointly optimize GS and PS to learn directly from the actual channel conditions instead of relying on the calculated channel gradient. By learning the actual channel conditions, GPS-AE is not affected by inaccurate channel modeling in actual communication systems, improving the E2E optimization efficiency of the AE neural network, and enabling the efficient deployment of the method of joint optimization of signal GS and PS in real communication systems.

[0035] The technical solution provided by this application will be described in detail below with reference to the accompanying drawings.

[0036] The following describes an end-to-end intelligent optimization method for a communication system without channel model assistance provided by an embodiment of the present application.

[0037] Figure 1 The basic architecture of the end-to-end intelligent communication system without channel model assistance is shown. The transmitter and receiver of the system contain trainable functional modules. The signal generated by the transmitter is transmitted through the real channel and then recovered at the receiver. The recovered information is compared with the transmitted information to calculate the recovery loss. Based on the loss, the trainable parameters of the transmitter and receiver are updated through the BP algorithm. The trainable parameters of the transmitter and receiver are represented by θ T and θ R Indicates. For example, at the transmitter, T-NN maps each m-bit information message into a complex constellation symbol x. Information message set By 2 m bit vector Each vector is composed of m information bits. T-NN uses the function Encode the N bit vector into a signal x = {x1, ..., x N}. This signal reaches the receiving end after undergoing modulation g(·) and real channel transmission h(·). Figure 1 The module 101 in the illustrated architecture is expanded, Figure 2 The process of performing channel estimation directly at the receiving end to achieve end-to-end optimization without the assistance of a channel model is shown. The core idea of ​​the model-free method proposed in this application is to use a trainable equalizer to estimate and equalize the channel impairments in the received signal. The total transfer function combining SCM modulation and the channel is expressed as The corresponding received signal is designated as First, calculate the difference between the Tx and Rx signals during actual transmission using the following formula:

[0038] Δ=vu.

[0039] Then we can simply express the computational graph between R-NN and T-NN in GPS-AE as:

[0040]

[0041] Assume that the inverse transfer function of SCM demodulation is expressed as g -1 (·), and the equalizer is expressed as Then the total inverse transfer function can be expressed as This embodiment provides an equalizer used in combination with SCM demodulation, which can perform the following operations:

[0042]

[0043] in represents the actual inverse transfer function, is the estimated inverse transfer function, and is the estimation error. If the neural equalizer can effectively reverse the channel effect to restore the signal, we can describe the equalization result as:

[0044]

[0045] In this embodiment, an end-to-end intelligent optimization method without the aid of a channel model is proposed. This embodiment uses a trainable equalizer to evaluate and subsequently mitigate channel impairments in the received signal. To simplify the channel impact on the transmitted signal u, it is formalized as an additive component Δ. When performing gradient calculation with v, Both Δ and Δ are considered constants, eliminating the need for derivatives. The channel information is embedded in the trainable equalizer, which is characterized by an inverse transfer function. Therefore, during the back propagation from the receiving neural network R-NN to the transmitting neural network T-NN, if an ideal equalizer If the training is sufficient, the channel derivatives will not be needed.

[0046] For example, a neural network-based trainable equalizer is trained by using the received signal v as input and the original symbol x as the label, using the mean square error (MSE) loss function, which is expressed as:

[0047]

[0048] Here, we use N pairs of equalized symbols and corresponding transmitted symbols in a batch to calculate the MSE, where i represents the index of different symbols in the batch. This MSE component is the total loss in the E2E optimization framework. After successful training, the equalizer outputs a signal similar to that of an AWGN channel dominated by additive noise. This equalized signal, affected by additive noise, can be efficiently demodulated and decoded by the receiver.

[0049] It should be noted that the trainable equalizer can be implemented using a neural network or other classic adaptive filters. The above-mentioned trainable equalizer is an example. The embodiments of the present application do not limit the specific structure or implementation method of the trainable equalizer. Any equalizer that can equalize the received signal and ensure that the equalized signal to be processed by the receiver is only affected by quasi-additive noise can be used to replace the trainable equalizer in the above example.

[0050] For example, when using a receiver neural network (R-NN) for bit metric decoding, the optimization process focuses on minimizing the binary cross entropy (BCE) loss function:

[0051]

[0052] in, It means to find the mathematical expectation. and Represents the received symbols The corresponding j-th bit information is b j The true probability and estimated probability of . This E2E optimization can be effectively implemented using the classic stochastic gradient descent (SGD) algorithm.

[0053] It should be noted that the above-mentioned bit-level information format is an example. The embodiment of the present application does not limit the format of the information to be encoded at the input of the transmitter and the format of the recovered information output by the receiver. Any original information format that can be used as the input of the neural network transmitter, including bits, symbols, and waveforms, can be encoded by T-NN and generate a transmission waveform. The signal after equalization by the trainable equalizer can be sent to the receiver in a form only interfered by additive noise to recover the original information.

[0054] It should also be noted that the receiver can be implemented through a neural network or other DSP algorithms that support propagation gradients, or a receiver that combines neural networks and DSP algorithms. The model-free end-to-end intelligent optimization method for the communication system proposed in the embodiment of the present application does not limit the scope of optimization of the transceiver. Any transceiver functional module that supports the transfer gradient can be used as an optimization target.

[0055] The following describes a method for joint shaping and optimization of signal constellation geometry and probability in an end-to-end intelligent communication system provided by an embodiment of the present application.

[0056] Figure 3 The signal constellation geometry and probability joint shaping optimization process in the end-to-end intelligent communication system is shown. Figure 1 The basic architecture of the end-to-end intelligent communication system without channel model assistance is similar to that shown in the figure. The transmitter needs to implement a more specific constellation PS and GS joint optimization task, which is also implemented using the end-to-end framework without channel model assistance. Figure 3 In the architecture shown, module 301 is expanded. Figure 4 Two constellation data generation methods used by the transmitter in the training phase and the deployment phase are shown. Figure 4 In module 401, during the training phase, the end-to-end intelligent communication system includes a transmitter neural network (T-NN) structure, which consists of a constellation probability generator responsible for PS, a sampler, an OBM, a BCM responsible for GS, and a subcarrier modulation (SCM) modulator. The constellation probability generator, OBM, and BCM are implemented using neural networks. The constellation probability generator outputs a probability vector Contains the PS configuration of each symbol. To ensure smooth propagation of gradients in T-NN, two gradient preservation methods are implemented. The first gradient preservation method uses the Gumbel-Softmax method. The Gumbel sampler generates one-hot encoding symbols consistent with the probability vector. The Gumbel-Softmax technique can keep the loss function relative to the probability vector. The gradient of Optimization. The T-NN transmitter omits a mapping mechanism during the training phase when using a sampler and OBM to generate PS symbol data. This mechanism can transform any original input bit stream so that the output symbols are independent and identically distributed (IID) and follow a given probability distribution. A practical example of this ideal mapping is obtained by applying the Constant Composition Distribution Matching (CCDM) technique.

[0057] See also Figure 4 In module 402, during the deployment phase, the CCDM mapper uses the probability vector generated by the constellation probability generator in the T-NN and the original payload bit data as input, generating a probability vector The distributed PS symbol data is converted into bit vectors by OBM. The mapper uses a second gradient-preserving method to train and fix the weights of a neural network to map one-hot inputs to soft bit vectors that approximate binary. A straight-through estimator is then used to obtain hard-decision bit vectors. During training, the bit vectors are used for forward propagation and the soft bit vectors for backward propagation.

[0058] The bit vector is input to the BCM mapper responsible for the GS, which is pre-trained to follow the standardized Gray mapping scheme for the standard grid-distributed Quadrature Amplitude Modulation (QAM) constellation. This pre-training scheme provides an advanced starting point for the bit-to-constellation mapper, accelerating convergence to the optimal GS configuration more effectively than random parameter initialization. The GS constellation is normalized according to its PS distribution as follows:

[0059]

[0060] in is the unnormalized GS constellation, which is the intermediate product of the bit-to-constellation mapper. The bit-to-constellation GS mapper outputs complex symbols x based on the input bit vectors, which follow the The normalized complex constellation symbols after PS and GS are sent to the modulator to generate the final transmit waveform (Tx).

[0061] It should be noted that the modulator can use any differentiable modulation method. Various basic modulation methods in communication systems, including single-carrier modulation and orthogonal frequency division multiplexing, are differentiable. The above-mentioned trainable equalizer is an example. The modulation format used by the modulator in the embodiment of the present application is limited. Any differentiable modulation method can be used to replace the subcarrier modulation method used by the modulator in the above example.

[0062] See also Figure 3 In the module 302 of the architecture shown, the end-to-end intelligent optimization method for communication systems without channel model assistance proposed in the present application is used. The transmitted waveform is transmitted through a real channel, and a trainable equalizer is used at the receiving end to perform channel estimation and equalization. After equalization, the signal is only affected by the residual additive noise. and the gradient of v, both and Δ are treated as constants, avoiding the need to differentiate them when backpropagating the gradient. The channel information is embedded in a trainable equalizer that is trained to implement the transfer function composed of the modulation and channel transfer. An inverse transfer function Therefore, in the reverse propagation from the receiver to the transmitter, if the ideal equalizer If the training is sufficient, the channel derivative is no longer needed. At the receiver, the main goal of bit-level decoding is to recover each corresponding received symbol y n The transmission symbol x n Corresponding bit label in

[0063] For example, the receiver estimates each bit The posterior probability The generalized mutual information (GMI) is used as a measure of the receiver decoding performance. GMI can be calculated according to the following formula:

[0064]

[0065] Where X and B j Respectively represent and x i and b (i) The j-th bit of the random variable is related to X, H(X) represents the information entropy of X. The parameter s is the optimization factor for error compensation, L j is the log-likelihood ratio (LLR) of the jth position

[0066]

[0067] The right side of the GMI calculation formula contains two terms: the entropy of the constellation and the decoding uncertainty, which reflects the channel noise effect.

[0068] For example, in order to fairly evaluate the performance of the model-free E2E optimization framework shown in this application, the receiver uses the mismatched Gaussian receiver (MGR) commonly used in classical communication systems to decode and output the recovered information. Under the premise that the noise in the constellation follows the AWGN model, GMI can be achieved by using the constellation geometry distribution obeyed by the transmitted constellation symbols. Probability distribution and the corresponding bit labels Monte Carlo simulation approximation yields:

[0069]

[0070] in Expressed as

[0071]

[0072] Here, and Represents the geometric distribution of constellations The i-th bit of its bit label has a value of a subset of 1 and 0.

[0073] It should be noted that the receiver can be implemented by a neural network, a mismatched Gaussian receiver (MGR), or other receivers that can output mutual information of demodulated signal bits during the execution of a DSP algorithm.

[0074] GPS-AE is composed of T-NN with PS and GS functions, real channel, trainable equalizer and receiver. In this embodiment, GMI is incorporated into the total loss after each forward propagation in the GPS-AE end-to-end training cycle. In. Through a normalized generalized mutual information (NGMI) constraint, the total loss function is expressed as:

[0075]

[0076] The NGMI is defined as:

[0077] NGMI=1-[H(X)-GMI] / m

[0078] Parameter NGMIt h Indicates the NGMI threshold that needs to be met during PS optimization to guarantee the expected bit rate of the Forward Error Correction (FEC) coding scheme. If the NGMI exceeds NGMI th , then the total loss function The third point will be cancelled, making Only the GMI and MSE contributions from the equalizer are included. Otherwise, if the NGMI does not reach the threshold, the third term comes into play in the loss function, which is used to compensate for NGMI below NGMIt h The loss function is calculated using a batch of N randomly sampled training symbols. After the gradient of T and θ R The stochastic gradient descent (SGD) method is used for updates. After each BP iteration, the GS and PS configurations are optimized to improve the performance of GMI.

[0079] The following describes the performance results of an end-to-end intelligent communication system based on the unassisted constellation geometry and probability joint shaping optimization technology provided by an embodiment of the present application under different signal baud rates and received optical powers (ROP) in actual optical fiber communications.

[0080] To verify the effectiveness of the embodiments of the present application, the performance of the end-to-end intelligent communication solution based on GPS-AE was compared with the MB distributed PS standard QAM signal baseline (BS) transmission solution with and without the neural network equalizer. Figure 5 (a) shows the changing trends of the Net Bitrate (NBR) and Achievable Information Rate (AIR) of the three schemes when adjusting the baud rate at 5dBm ROP. Given that the bandwidth limit of the transmission system is 59GHz, all configurations show performance degradation when the baud rate exceeds 60GBaud. GPS-AE reaches peak performance at 60Gbaud, with an AIR of 366.9Gb / s and an NBR of 341.2Gb / s, which means that the bandwidth limit is effectively managed through the synergy of the jointly trained neural equalizer and T-NN. Figure 5 (b) The performance improvement of GPS-AE is recorded, reaching an AIR of 373.0Gb / s and an NBR of 349.2Gb / s when the ROP is increased to 7dBm. It is worth noting that the NBR of the baseline transmission scheme without neural network equalization lags behind GPS-AE by 51.4Gb / s at 5dBm ROP. In addition, when comparing the transmission case without neural network, the baseline transmission scheme without neural network is more susceptible to nonlinear distortion, while GPS-AE is able to flexibly optimize the transmitted signal to cope with dynamic channel changes when the ROP increases. This study symbolizes the practical value of end-to-end optimization techniques without channel model assistance in significantly enhancing the practical value of PS and GS joint signal optimization methods, provides strong experimental evidence to support its practical applicability and efficiency, and bridges the challenging gap between theoretical construction and its operational deployment.

[0081] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, devices and systems. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices and systems according to the embodiments of the present application. The disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these changes and variations.

[0082] It should be understood that the specific examples in the embodiments of this application are only to help those skilled in the art better understand the technical solutions of this application. The above specific implementation methods can be considered as the optimal implementation methods of this application, rather than limiting the scope of the embodiments of this application.

[0083] Those skilled in the art will appreciate that the algorithmic steps of each example described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, and such implementation should not be considered beyond the scope of this application.

[0084] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for optimizing the combined geometry and probability of transmitted signal constellation shaping in an end-to-end intelligent communication system, characterized in that The method is applied to the optimization design of a transmission signal of a communication system, and the method comprises: Use end-to-end optimization methods to perform intelligent optimization without the assistance of channel models; The autoencoder neural network of the end-to-end intelligent communication system includes a neural network transmitter, a real channel, a trainable equalizer and a receiver; The neural network transmitter is designed in a modular way, including: The modular neural network transmitter includes a constellation probability generator, a sampler, a one-hot code to bit vector mapper, a bit vector to constellation geometry mapper, and a modulator. The constellation probability generator is used to output a probability vector, which contains the probability configuration of each constellation point; The sampler generates symbols using one-hot encoding based on probability vector sampling, and these symbols correspond one-to-one to constellation points in the constellation diagram to achieve probabilistic shaping of the constellation; The one-hot encoded symbols are converted into bit vectors by a one-hot code to bit vector mapper, and different bit vectors are input into a bit vector to constellation geometric position mapper and mapped to constellation points at different positions to achieve geometric shaping of the constellation points; The probabilistically and geometrically shaped symbols are fed into a modulator to generate the final transmitted waveform. This waveform is then transmitted through a real channel and, at the receiver, a trainable equalizer is used for channel estimation and equalization. The equalized signal is then affected by residual additive noise. After demodulation and decoding at the receiver, the error between the recovered bits at the receiver and the original transmitted bits is calculated. The end-to-end optimization method includes: A neural network is used to construct a communication system transmitter. The encoded signal is transmitted through a real channel and then reaches the receiver. The received signal first passes through a trainable equalizer connected to the front end of the receiver to perform channel estimation and equalization. The equalized signal is affected by quasi-additive noise and is input into the receiver for demodulation, decoding and recovery of the original information; The receiver is implemented using a neural network or a digital signal processing algorithm; the end-to-end optimization framework of the autoencoder neural network includes the transmitter, a real channel, a trainable equalizer and a receiver; the equalizer, transmitter and receiver in the end-to-end optimization framework are trained.

2. The method for optimizing the transmission signal constellation geometry and probability in an end-to-end intelligent communication system according to claim 1, characterized in that: The equalizer, transmitter, and receiver are trained simultaneously in an end-to-end optimization framework. The training process includes: The total loss function of the end-to-end optimization framework includes the error of the equalizer in channel estimation and equalization and the error after receiver decoding; the backpropagation algorithm updates the parameters of each trainable part in the end-to-end optimization framework based on the gradient of the total loss function, so that the updated transmitter can achieve improved coding, the updated trainable equalizer can achieve better channel estimation and equalization, and the receiver can achieve improved decoding of the equalized signal.

3. The method for optimizing the transmission signal constellation geometry and probability in an end-to-end intelligent communication system according to claim 2, characterized in that: Supports simultaneous communication and end-to-end optimized transmission data structures in end-to-end communication systems, including: The transmission data generated by the transmitter includes preamble training data and payload; The leading training data participates in the calculation of the total loss function and updates the parameters of the trainable part of the autoencoder of the end-to-end optimization framework through the back-propagation algorithm; The payload is used to calculate the receiver decoding error rate in the initialization phase of the end-to-end communication system to characterize the training effect and trigger the training termination condition, and directly carries the communication data for end-to-end communication in the deployment and operation phase of the end-to-end communication system; The proportion of the training data and payload in the transmitted data is dynamically adjusted based on the performance of the end-to-end communication system on the payload. When the payload can be recovered at a lower bit error rate at the receiving end, the proportion of the training data in the transmitted data is reduced; when the decoding error of the payload at the receiving end is large, the proportion of the training data in the transmitted data is increased, thereby strengthening the training of the autoencoding neural network to reduce the decoding error of future received data.

4. The method for joint shaping optimization of transmission signal constellation geometry and probability in an end-to-end intelligent communication system according to claim 1, characterized in that: The bit vector to constellation geometric position mapper includes: The bit vector to constellation geometry position mapper is pre-trained using a standard bit to lattice constellation Gray mapping relationship before being embedded in the autoencoder neural network; Geometric shaping is implemented using a mapper that maps the bit vector to the geometric position of the constellation. This is achieved through a neural network whose input is the bit vector and output is the position of the constellation point. While adjusting the geometric positions of the constellation points, the bit labels corresponding to the constellation points are also optimized, ensuring that when demapping the constellation of the received signal, misjudgment of the constellation points leads to the least bit errors.

5. The method for joint shaping optimization of transmission signal constellation geometry and probability in an end-to-end intelligent communication system according to claim 1, characterized in that: The loss function of the autoencoder neural network for the joint shaping optimization method of the transmitted signal constellation geometry and probability in the end-to-end intelligent communication system includes: The trainable equalizer performs channel estimation and equalization, and the equalization error between the equalized signal and the transmitted signal; After the demodulation and decoding operations at the receiver, the recovery error between the bit information recovered at the receiving end and the original transmitted bit information is calculated; The recovery error and the equalizer's equalization error together constitute the total loss function of the autoencoder neural network. Based on this total loss function, the backpropagation algorithm updates the parameters of the trainable constellation probability generator, the bit vector to constellation geometry position mapper, and the equalizer in the end-to-end intelligent communication system, thereby achieving joint optimization of the communication system's transceiver.

6. The method for joint shaping optimization of transmission signal constellation geometry and probability in an end-to-end intelligent communication system according to claim 1, characterized in that Transmitter module using two gradient preservation methods, including: The first gradient-preserving method ensures that the gradient can be propagated throughout the entire process from the constellation probability generator generating the probability vector to the final waveform generation in the transmitter. The sampler uses the Gumbel-Softmax method to generate one-hot encoded symbols consistent with the probability vector through the Gumbel sampler and ensures the gradient of the total loss function with respect to the probability vector. In the second gradient-preserving method, the one-hot encoded symbols are converted into bit vectors by a pre-trained and fixed one-hot code to bit vector mapper. The one-hot code to bit vector mapper implemented by a neural network ensures that it can be embedded in the transmitter for gradient propagation.

7. A communication device based on a neural network, characterized in that: The method according to any one of claims 1 to 6 is adopted.

Images