A deep learning-based geometric constellation shaping joint coding modulation method

Abstract

The application discloses a geometric constellation shaping joint coding modulation method based on deep learning, and belongs to the geometric constellation shaping method. By constructing an end-to-end system containing a deep learning modulator and a demodulator, a log-likelihood ratio calibration module is integrated at the end of the demodulator to output soft information matched with a decoder; a multi-objective joint loss function containing a cross-entropy loss, a generalized mutual information loss and a boundary constraint loss is determined; a phased training strategy of first freezing the calibration module for basic constellation shaping and then jointly fine-tuning is adopted; the trained modulator is deployed at a sending end for geometric constellation shaping mapping, and the demodulator is deployed at a receiving end to output a calibrated log-likelihood ratio sequence to the decoder. The application realizes deep collaboration of coding and modulation, and improves transmission reliability and spectral efficiency.

Description

Technical Field

[0001] This invention relates to a geometric constellation shaping method, and more particularly to a deep learning-based geometric constellation shaping joint coding modulation method, apparatus, device, and storage medium. Background Technology

[0002] To meet the unprecedented demands of 6G communication systems for massive data transmission, wireless communication must continuously improve its channel capacity and transmission reliability. However, the orthogonal amplitude modulation constellations widely used in traditional communication systems are increasingly becoming performance bottlenecks due to their fixed geometric structures. Standard constellation modulation schemes designed based on idealized channel models struggle to adapt to complex real-world channel environments and diverse service objectives. Especially in higher-order modulation, their fixed dot spacing leads to a significant deterioration in bit error rate performance, completely hindering further approaches to the Shannon limit.

[0003] With the development of artificial intelligence and end-to-end learning technologies, intelligent methods to address the bottlenecks of traditional constellation shaping are being explored. In existing research, some works are based on numerical optimization algorithms for specific channels, which can search for high-performance constellation diagrams under a given channel model. However, their optimization results heavily depend on the accuracy of the preset channel model and lack the ability to generalize to unknown or time-varying channels.

[0004] Another area of ​​research focuses on lookup table-based adaptive modulation and coding schemes, which adapt to channel variations by switching between a finite number of pre-defined constellation diagrams and coding rate combinations. However, this approach is inherently discrete and discontinuous, with a limited range of choices, making it impossible to achieve broad performance optimization. In recent years, autoencoder-based methods have received widespread attention and research as a flexible, data-driven end-to-end optimization paradigm. The core idea is to construct the communication system as an end-to-end neural network, enabling the constellation diagram to be learned autonomously through a data-driven training process, rather than relying on manual pre-setting.

[0005] However, directly applying autoencoder-based constellation shaping methods to practical wireless communication systems can lead to difficulties in adapting to the encoded bitstream, resulting in decreased end-to-end transmission reliability. Specifically, constellation shaping methods based on intermediate index optimization struggle to consider the characteristics of the channel encoder, thus failing to adaptively adjust the joint error correction strategy according to actual channel conditions. In particular, the core objective of wireless communication scenarios is to achieve efficient and highly reliable data transmission, often demanding extremely high bit error rate performance. Therefore, while leveraging the flexible and adaptive advantages of constellation shaping, minimizing the negative impacts of separating and optimizing coding and modulation, and achieving deep synergy and joint optimization of channel coding and constellation modulation, has become a critical challenge that urgently needs to be addressed. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] To address this, this invention proposes a deep learning-based geometric constellation shaping and joint coding modulation method. By constructing an end-to-end communication system model comprising a deep learning modulator and a demodulator with an integrated log-likelihood ratio calibration module, the original log-likelihood ratio is calibrated to a log-likelihood ratio sequence matching the channel decoder. A multi-objective joint loss function, including cross-entropy loss, generalized mutual information loss, and boundary constraint loss, is determined. A phased training strategy is adopted, first freezing the calibration module for basic constellation shaping, then performing end-to-end joint fine-tuning. The trained modulator is deployed at the transmitting end for geometric constellation shaping mapping and transmission, and the trained demodulator is deployed at the receiving end to output the calibrated log-likelihood ratio sequence to the decoder. This achieves deep collaboration between coding and modulation, effectively improving the transmission reliability and spectral efficiency of the wireless communication system.

[0008] Another objective of this invention is to propose a deep learning-based geometric constellation shaping joint coding modulation device.

[0009] The third objective of this invention is to provide a computer device.

[0010] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0011] To achieve the above objectives, this invention proposes a deep learning-based geometric constellation shaping joint coding and modulation method, comprising:

[0012] Obtain the initial communication system model and configure it to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. Configure the log-likelihood ratio calibration module to calibrate the raw log-likelihood output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder. Output the completed end-to-end communication system model. Obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function; Using the end-to-end communication system model and the multi-objective joint loss function, a phased joint training strategy is executed. In the first phase, the parameters of the log-likelihood ratio calibration module are frozen, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freeze of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output. The trained deep learning modulator parameters are deployed at the transmitting end, and the deep learning modulator is used to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted. The resulting constellation symbols are then sent to the receiving end via a wireless channel. The trained deep learning demodulator parameters are deployed at the receiver, and the received signal is demodulated using the deep learning demodulator. The log-likelihood ratio calibration module outputs the calibrated log-likelihood ratio sequence, and the log-likelihood ratio sequence is output to the channel decoder for decoding, outputting the recovered user information bits.

[0013] In one embodiment of the present invention, constructing an end-to-end communication system model includes: Construct a deep learning modulator, which includes an input layer, a feature extraction module, and a normalization layer; among which, The input layer uses a fully connected layer to map the grouped one-hot encoded vectors into a high-dimensional feature tensor; The feature extraction module consists of multiple cascaded residual blocks. Each residual block contains two fully connected layers and an activation function, and has skip connections from input to output. The normalization layer maps the high-dimensional features output by the feature extraction module into two-dimensional vectors, corresponding to the in-phase components and the orthogonal components respectively, and normalizes the power of the output constellation symbols to limit the average power of the transmitted symbols. A deep learning demodulator is constructed, comprising a basic demodulation module and a log-likelihood ratio calibration module; wherein... The basic demodulation module is symmetrically configured with the deep learning modulator, and includes an input layer that maps the received signal into high-dimensional features and multiple residual blocks for feature refinement, which output the uncalibrated raw log odds. The log-likelihood ratio calibration module is located at the end of the basic demodulation module. It consists of an affine transformation layer with independently configurable trainable scaling and bias parameters for each bit position. It is used to perform bit-by-bit adaptive calibration of the original log-likelihood ratio and output a log-likelihood ratio sequence that matches the input distribution of the channel decoder.

[0014] In one embodiment of the present invention, determining the multi-objective joint loss function includes: The bit error rate loss term is determined by using a binary cross-entropy loss function to measure the difference between the bit prediction probability of the demodulator output and the actual information bit tag at the transmitter. A generalized mutual information loss term is determined and constructed based on the lower bound formula of mutual information. It is used to characterize the channel capacity of the communication system under the current signal-to-noise ratio and constellation shape. This term is optimized to promote the distribution of constellation points in the direction of maximum channel capacity. The log-likelihood ratio boundary constraint loss term is determined, and a threshold-based regularization term is introduced. When the absolute value of the log-likelihood ratio output by the demodulator is less than a preset threshold, a penalty value is generated to constrain the confidence of the log-likelihood ratio. The multi-objective joint loss function is obtained by weighting and summing the bit error rate loss term, the generalized mutual information loss term, and the log-likelihood ratio boundary constraint loss term. Each term is configured with a non-negative hyperparameter weight to balance bit error rate performance and channel capacity performance.

[0015] In one embodiment of the present invention, a phased joint training strategy is implemented, including: The first phase of training is performed by setting the parameters of the log-likelihood ratio calibration module to an untrainable state. The backpropagation algorithm only updates the weights of the deep learning modulator and the basic demodulation module to perform basic constellation shaping without log-likelihood ratio calibration interference. After the first stage of training is completed, the second stage of training is performed to remove the untrainable state of the parameters of the log-likelihood ratio calibration module, so that the log-likelihood ratio calibration module can participate in joint training and perform end-to-end fine-tuning on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module, so as to achieve joint optimization and adaptation of modulation and coding. When the joint loss function reaches the preset convergence condition or the training rounds reach the preset threshold, the trained modulator and demodulator parameters are saved.

[0016] In one embodiment of the present invention, a trained deep learning modulator is deployed at the transmitting end, and the encoded bitstream to be transmitted is geometrically constellated shaped and mapped before transmission, including: The transmitting device loads the trained modulator parameters and deploys the network weights of the deep learning modulator obtained from offline training to the baseband processing module of the transmitter. Channel coding is performed on the user data to be transmitted, and the encoded bit stream is divided into bit groups according to the modulation order; The divided bit groups are input into the deep learning modulator, which maps each bit group into a constellation symbol after geometric constellation shaping according to the trained parameters. The constellation symbol presents a non-uniformly distributed shape on the complex plane, and the distance between different constellation points is adaptively adjusted according to the protection requirements of the bits they carry. The generated constellation symbols are transmitted to the receiving end via a wireless channel.

[0017] In one embodiment of the present invention, a trained deep learning demodulator is deployed at the receiving end to demodulate the received signal. A calibrated log-likelihood ratio sequence is output from the log-likelihood ratio calibration module to the channel decoder for decoding, including: The receiving device loads the trained demodulator parameters and deploys the network weights of the deep learning demodulator obtained from offline training to the signal processing module of the receiver. The network weights include the trainable scaling parameters and bias parameters of the log-likelihood ratio calibration module. The received signal is subjected to channel estimation and equalization processing. The channel state information is estimated using the pilot signal, and the equalization algorithm is used to eliminate the phase rotation and amplitude distortion caused by channel fading, so as to obtain the equalized signal. The equalized signal is input into the deep learning demodulator. After feature extraction by the basic demodulation module, the log-likelihood ratio calibration module performs affine transformation on the extracted raw log-likelihood ratio and directly outputs the log-likelihood ratio sequence for each encoded bit. The log-likelihood ratio sequence is input into the channel decoder, which uses the belief propagation algorithm to perform iterative decoding to recover the original user information bits.

[0018] This invention also proposes a deep learning-based geometric constellation shaping joint coding and modulation device, comprising: The joint coding and modulation system construction module is used to obtain the initial communication system model. The initial communication system model is configured to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. The log-likelihood ratio calibration module is configured to calibrate the raw log-probability output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder. The module outputs the completed end-to-end communication system model. The multi-objective joint loss function construction module is used to obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function. The phased joint training strategy execution module is used to execute a phased joint training strategy using the end-to-end communication system model and the multi-objective joint loss function. In the first phase, the parameters of the log-likelihood ratio calibration module are frozen, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freezing state of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output. The deep learning modulator deployment and mapping transmission module is used to deploy the trained deep learning modulator parameters at the transmitting end, use the deep learning modulator to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted, and send the mapped constellation symbols to the receiving end via a wireless channel. The deep learning demodulator deployment and LLR decoding module is used to deploy the trained deep learning demodulator parameters at the receiver, demodulate the received signal using the deep learning demodulator, output the calibrated log-likelihood ratio sequence from the log-likelihood ratio calibration module, and output the log-likelihood ratio sequence to the channel decoder for decoding, outputting the recovered user information bits.

[0019] This invention discloses a deep learning-based geometric constellation shaping and joint coding-modulation method. By constructing an end-to-end communication system comprising a deep learning modulator and an integrated log-likelihood ratio calibration module, and employing a multi-objective joint loss function including cross-entropy loss, generalized mutual information loss, and boundary constraint loss, coupled with a staged training strategy of first freezing the calibration module and then jointly fine-tuning, this method effectively solves the problems of redundant conversion, numerical accuracy loss, and code-modulation separation optimization caused by interface mismatch between the demodulator and decoder in existing technologies. It achieves integrated joint optimization from constellation shaping and soft information calibration to channel decoding, significantly improving the transmission reliability and spectral efficiency of the wireless communication system, and enhancing the model's generalization ability and engineering applicability under complex channel conditions.

[0020] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing a deep learning-based geometric constellation shaping joint coding modulation method as described in the first aspect embodiment.

[0021] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a deep learning-based geometric constellation shaping joint coding modulation method as described in the first aspect embodiment.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the overall process of the deep learning-based geometric constellation shaping joint coding and modulation method described in this invention; Figure 2 This is a schematic diagram of the network architecture of the deep learning modulator and demodulator provided for the deep learning-based geometric constellation shaping joint coding modulation method described in this invention. Figure 3 This is a schematic diagram of the two-stage joint training process for the deep learning-based geometric constellation shaping joint coding and modulation method described in this invention. Figure 4 This is a schematic diagram of the data transmission and LLR recovery process provided by the deep learning-based geometric constellation shaping joint coding and modulation method of the present invention. Figure 5 This is a schematic diagram of the structure of a deep learning-based geometric constellation shaping and joint coding modulation device according to an embodiment of the present invention. Figure 6 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] The following description, with reference to the accompanying drawings, describes a deep learning-based geometric constellation shaping joint coding modulation method and apparatus according to an embodiment of the present invention.

[0027] A deep learning-based geometric constellation shaping joint coding and modulation method includes the following steps: S1. Obtain the initial communication system model and configure it to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. Configure the log-likelihood ratio calibration module to calibrate the raw log-probability output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder. Output the completed end-to-end communication system model. S2, obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function; S3, using the end-to-end communication system model and the multi-objective joint loss function, a phased joint training strategy is executed. In the first phase, the parameters of the log-likelihood ratio calibration module are set to a frozen state, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freezing state of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output. S4, deploy the trained deep learning modulator parameters at the transmitting end, use the deep learning modulator to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted, and send the mapped constellation symbols to the receiving end via the wireless channel. S5, deploy the trained deep learning demodulator parameters at the receiver, use the deep learning demodulator to demodulate the received signal, the log-likelihood ratio calibration module outputs the calibrated log-likelihood ratio sequence, the log-likelihood ratio sequence is output to the channel decoder for decoding, and the recovered user information bits are output.

[0028] Specifically, such as Figure 1 As shown, the deep learning-based geometric constellation shaping joint coding and modulation method of the present invention includes: Step 1: Construct a training dataset for a joint coding-modulation communication system; Step 2: Construct a transceiver neural network model based on a residual network and an LLR calibrator; Step 3: Determine the joint loss function that includes mutual information and boundary constraints; Step 4: To address the issue of convergence difficulties in joint optimization, a phased joint training strategy is implemented. Step 5: The sending end loads the optimal parameters, performs geometric constellation shaping mapping, and sends the data; Step 6: The receiver performs online LLR calculation and decoding based on the channel equalization signal.

[0029] Furthermore, in some specific steps of the present invention, step 1, which involves constructing a training dataset for a joint coding-modulation communication system, specifically includes: Step 101: The system generates a random original information bit sequence. The system simulates a source to generate a binary random bit stream sequence. The length K of the information bit sequence is set according to the actual communication standard, such as the 5G NR standard. Step 102: Channel coding is performed on the information bit sequence. A low-density parity-check code (LDPC) encoder is used to convert the shorter information bit sequence into a longer coded bit sequence. The code rate is set to R=1 / 2 and the length of the encoded sequence is N=1944. This coding process introduces redundant information, which enables the subsequent neural network model to learn the intrinsic relationship between the coding structure and the modulation symbols. Step 103: Divide the generated encoded bit sequence into a training dataset and a test dataset, which are used for parameter optimization and performance evaluation of the model, respectively.

[0030] Furthermore, in some specific steps of the present invention, such as Figure 2 As shown, Figure 2 The diagram shows a network architecture of a deep learning modulator and demodulator. Step 2, which involves constructing a transceiver neural network model based on a residual network and an LLR calibrator, specifically includes: Step 201: Construct a deep learning modulator to map the input bit group into constellation symbols on a two-dimensional complex plane. The input of the modulator is a grouped one-hot encoded vector, and the output is a two-dimensional complex symbol. The modulator includes an input layer, a feature extraction module, and a normalization layer. The input layer uses a fully connected layer to map the one-hot vector of length M into a high-dimensional feature tensor to capture the nonlinear features between bits in a higher-dimensional space. The feature extraction module consists of multiple cascaded residual blocks. Each residual block contains two fully connected layers and a ReLU activation function, and a skip connection is set from input to output. This residual structure can effectively solve the gradient vanishing problem in deep network training and ensure that the modulator can learn more complex constellation geometric distributions. The normalization layer maps the high-dimensional features into two-dimensional vectors corresponding to the in-phase component I and the orthogonal component Q, respectively, and introduces power normalization. By dividing by the root mean square value of the symbols in the current batch, the average power of the transmitted symbols is strictly limited to 1. Step 202: Construct a deep learning demodulator. The input to the demodulator is the equalized received signal, and the output is the log-likelihood ratio (LLR) for each bit. The demodulator comprises a basic demodulation module and an LLR calibration module. The basic demodulation module, used for initial feature extraction, has a symmetrical structure to the modulator and also includes an input layer that maps the received signal to high-dimensional features, as well as three residual blocks for feature refinement. This module outputs the uncalibrated raw log-likelihood ratio. The LLR calibration module, located at the end of the network, consists of an affine transform layer. This module configures trainable scaling and bias parameters for each bit position to calibrate the log-likelihood ratio (LLR) of the demodulated output to match the input requirements of the channel decoder. This design allows the neural network to adaptively compensate for reliability differences at different bit positions, outputting a high-quality LLR consistent with the input distribution of the LDPC decoder.

[0031] Furthermore, in some specific steps of this invention, step 3, which involves determining a joint loss function that includes mutual information and boundary constraints, defines a multi-objective joint loss function to guide the optimization of model parameters in order to train the aforementioned model. Specifically, this includes: Step 301: Determine the bit error rate loss term using the binary cross-entropy loss function, which directly measures the distance between the bit probability output by the model and the true label. This is a core indicator for ensuring the basic reliability of the communication system, as shown in the following formula:

[0032] In the above formula (1), This represents the bit error rate loss term; This represents the total number of bits in a single training batch. Indicates the first... A real information bit tag; Indicates the output of the demodulator. Step 302 determines the generalized mutual information loss term by calculating the probability that each bit is predicted to be 1.

[0033] The loss term is constructed based on the lower bound formula of mutual information. This metric reflects the maximum bit rate that the system can theoretically transmit without errors under the current signal-to-noise ratio and constellation shape. Optimizing this term can promote the distribution of constellation points in the direction of maximum channel capacity. (2) In the above formula (2), Represents the generalized mutual information loss term; The constellation symbol representing the sending end; This represents the signal after equalization at the receiving end; Indicates a given transmission symbol When the signal is received Channel transition probability density; Indicates received signal Marginal probability density: This indicates the operation of calculating the mathematical expectation for all transmit and receive symbol pairs.

[0034] Step 303: Determine the LLR boundary constraint loss term and introduce a threshold-based regularization term. When the absolute value of the LLR output by the model is less than a preset threshold, a penalty value is generated to prevent the network from outputting uncertain soft information, thereby improving the confidence of the demodulation results. The penalty calculation formula is shown in the following formula (3): (3) In the above formula (3), This represents the LLR boundary constraint loss term; Indicates the total length of the bit sequence; Indicates the output of the demodulator. The log-likelihood ratio of each encoded bit; This indicates the preset hard confidence threshold. The operator indicates when the absolute value of LLR is below the threshold. A linear penalty is applied when the penalty is applied; otherwise, no penalty is applied.

[0035] Step 304: The three losses mentioned above are weighted and summed to obtain the final joint loss function. By adjusting the weighting coefficients, the bit error rate performance and channel capacity performance can be flexibly balanced. Step 304: The three losses mentioned above are weighted and summed to obtain the final joint loss function. By adjusting the weight coefficients, the bit error rate performance and channel capacity performance are flexibly balanced. The overall joint loss function is shown in equation (4) below: (4) In the above formula (4), The joint loss function representing the entire system; , , These represent the bit error rate loss terms. Generalized mutual information loss item and LLR boundary constraint loss term Non-negative hyperparameter weights.

[0036] Furthermore, in some specific steps of the present invention, such as Figure 3 As shown, Figure 3 The diagram illustrates a two-stage joint training process to address the convergence problem of joint optimization. Step 4, which involves implementing a phased joint training strategy, specifically includes: Step 401, perform the first stage of training: basic constellation coarse shaping. In this stage, at the beginning of training, the parameters of the LLR calibration module in the demodulator are set to an untrainable state. At this time, the backpropagation algorithm only updates the weights of the modulator and the basic demodulation module. The purpose of this stage is to allow the modulator to quickly explore a roughly reasonable constellation geometry without LLR calibration interference, so as to avoid gradient explosion or network non-convergence caused by the randomness of calibration parameter initialization. Step 402, perform the second stage of training: joint fine calibration. After the first stage is completed, unlock the LLR calibration module and let it participate in the joint training. At this time, the total loss is calculated using the accurate LLR value after calibration, and the entire communication link including the modulator, demodulator and calibrator is finely tuned end-to-end. In this stage, the position of the constellation point will be slightly offset according to the feedback of LLR, and the calibration parameters will converge to the optimal value, thereby achieving perfect adaptation of modulation and coding. Step 403: Save the optimal model parameters. When the loss function no longer decreases or reaches the preset training rounds, save the trained modulator and demodulator parameters.

[0037] Furthermore, in some specific steps of the present invention, such as Figure 4 As shown, Figure 4 This is a schematic diagram of the online data transmission and LLR recovery process provided in this embodiment. Step 5, where the sending end loads the optimal parameters and performs geometric constellation shaping mapping and transmission, includes the following steps: Step 501: The transmitting device loads the trained modulator parameters and deploys the neural network weights obtained from offline training to the baseband processing of the transmitter. Step 502: The transmitting end encodes and groups the user data to be transmitted. After the real-time bit stream is encoded by LDPC, each m bits are divided into a group. For example, for 64QAM modulation, each group consists of 6 bits. Step 503: The grouped data is input into the deep learning modulator. The modulator maps the data directly into geometrically shaped constellation symbols according to the trained parameters. The constellation diagram presents a non-uniformly distributed shape, rather than a traditional regular rectangular shape, and the distance between different constellation points is automatically adjusted according to the protection requirements of the bits they carry. Step 504: The generated constellation symbol is sent to the receiving end via a wireless channel. During the transmission process, the wireless channel will experience additive white Gaussian noise (AWGN) or Rayleigh fading.

[0038] Furthermore, in some specific steps of the present invention, step 6, in which the receiving end performs online LLR calculation and decoding based on the channel equalization signal, specifically includes: Step 601: The receiving device loads the trained demodulator parameters and deploys the weights obtained from offline training, including key calibration parameters, into the signal processing module of the receiver. Step 602: The receiver performs channel estimation and equalization on the received signal to eliminate some channel interference. The receiver performs channel estimation and equalization on the received signal, uses pilot signals to estimate channel state information (CSI), and uses the minimum mean square error (MMSE) equalization algorithm to eliminate phase rotation and amplitude distortion caused by channel fading; Step 603: The equalized signal is input into the deep learning demodulator. The demodulator performs feature transformation on the signal and passes it through the affine transformation processing of the LLR calibration module. It directly outputs a high-precision LLR sequence for each coded bit. Compared with the traditional method, the LLR sequence reflects the reliability probability of the bit more accurately. Step 604: Input the LLR sequence into a standard LDPC decoder. The decoder uses the belief propagation algorithm (BP) to iteratively decode the high-quality LLR, and finally recovers the original user information bits, thus completing the highly reliable data transmission.

[0039] The method of this invention effectively solves the problems of redundant conversion, numerical accuracy loss and coding-modulation separation optimization caused by the mismatch between the demodulator and decoder interfaces in the prior art. It realizes integrated joint optimization from constellation shaping, soft information calibration to channel decoding, significantly improves the transmission reliability and spectral efficiency of wireless communication systems, and enhances the generalization ability and engineering applicability of the model under complex channel conditions.

[0040] To implement the method of the above embodiments, such as Figure 5 As shown, this invention proposes a deep learning-based geometric constellation shaping joint coding and modulation device 10, comprising: The joint coding and modulation system construction module 100 is used to obtain an initial communication system model, which is configured to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. The log-likelihood ratio calibration module is configured to calibrate the raw log-probability output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder, and outputs the completed end-to-end communication system model.

[0041] The multi-objective joint loss function construction module 200 is used to obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function.

[0042] The phased joint training strategy execution module 300 is used to execute a phased joint training strategy using the end-to-end communication system model and the multi-objective joint loss function. In the first phase, the parameters of the log-likelihood ratio calibration module are frozen, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freeze of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output.

[0043] The deep learning modulator deployment and mapping transmission module 400 is used to deploy the trained deep learning modulator parameters at the transmitting end, use the deep learning modulator to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted, and send the mapped constellation symbols to the receiving end via a wireless channel.

[0044] The deep learning demodulator deployment and LLR decoding module 500 is used to deploy the trained deep learning demodulator parameters at the receiver, demodulate the received signal using the deep learning demodulator, output the calibrated log-likelihood ratio sequence from the log-likelihood ratio calibration module, and output the log-likelihood ratio sequence to the channel decoder for decoding, outputting the recovered user information bits.

[0045] The device of this invention effectively solves the problems of redundant conversion, accuracy loss and coding-modulation separation optimization caused by interface mismatch in the prior art. It realizes an integrated architecture from model building, joint training to online deployment, significantly improves the transmission reliability and spectrum efficiency of wireless communication systems, and enhances the engineering applicability of the device in complex channel environments.

[0046] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 6 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the deep learning-based geometric constellation shaping joint coding modulation method described above.

[0047] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0048] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A deep learning-based geometric constellation shaping joint coding and modulation method, characterized in that, include: Obtain the initial communication system model and configure it to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. Configure the log-likelihood ratio calibration module to calibrate the raw log-likelihood output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder. Output the completed end-to-end communication system model. Obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function; Using the end-to-end communication system model and the multi-objective joint loss function, a phased joint training strategy is executed. In the first phase, the parameters of the log-likelihood ratio calibration module are frozen, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freeze of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output. The trained deep learning modulator parameters are deployed at the transmitting end, and the deep learning modulator is used to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted. The resulting constellation symbols are then sent to the receiving end via a wireless channel. The trained deep learning demodulator parameters are deployed at the receiver, and the received signal is demodulated using the deep learning demodulator. The log-likelihood ratio calibration module outputs the calibrated log-likelihood ratio sequence, and the log-likelihood ratio sequence is output to the channel decoder for decoding, outputting the recovered user information bits.

2. The method as described in claim 1, characterized in that, Constructing an end-to-end communication system model includes: Construct a deep learning modulator, which includes an input layer, a feature extraction module, and a normalization layer; among which, The input layer uses a fully connected layer to map the grouped one-hot encoded vectors into a high-dimensional feature tensor; The feature extraction module consists of multiple cascaded residual blocks. Each residual block contains two fully connected layers and an activation function, and has skip connections from input to output. The normalization layer maps the high-dimensional features output by the feature extraction module into two-dimensional vectors, corresponding to the in-phase components and the orthogonal components respectively, and normalizes the power of the output constellation symbols to limit the average power of the transmitted symbols. A deep learning demodulator is constructed, comprising a basic demodulation module and a log-likelihood ratio calibration module; wherein... The basic demodulation module is symmetrically configured with the deep learning modulator, and includes an input layer that maps the received signal into high-dimensional features and multiple residual blocks for feature refinement, which output the uncalibrated raw log odds. The log-likelihood ratio calibration module is located at the end of the basic demodulation module. It consists of an affine transformation layer with independently configurable trainable scaling and bias parameters for each bit position. It is used to perform bit-by-bit adaptive calibration of the original log-likelihood ratio and output a log-likelihood ratio sequence that matches the input distribution of the channel decoder.

3. The method as described in claim 1, characterized in that, Determine the joint loss function for multiple objectives, including: The bit error rate loss term is determined by using a binary cross-entropy loss function to measure the difference between the bit prediction probability of the demodulator output and the actual information bit tag at the transmitter. A generalized mutual information loss term is determined and constructed based on the lower bound formula of mutual information. It is used to characterize the channel capacity of the communication system under the current signal-to-noise ratio and constellation shape. This term is optimized to promote the distribution of constellation points in the direction of maximum channel capacity. The log-likelihood ratio boundary constraint loss term is determined, and a threshold-based regularization term is introduced. When the absolute value of the log-likelihood ratio output by the demodulator is less than a preset threshold, a penalty value is generated to constrain the confidence of the log-likelihood ratio. The multi-objective joint loss function is obtained by weighting and summing the bit error rate loss term, the generalized mutual information loss term, and the log-likelihood ratio boundary constraint loss term. Each term is configured with a non-negative hyperparameter weight to balance bit error rate performance and channel capacity performance.

4. The method as described in claim 1, characterized in that, Implement a phased joint training strategy, including: The first phase of training is performed by setting the parameters of the log-likelihood ratio calibration module to an untrainable state. The backpropagation algorithm only updates the weights of the deep learning modulator and the basic demodulation module to perform basic constellation shaping without log-likelihood ratio calibration interference. After the first stage of training is completed, the second stage of training is performed to remove the untrainable state of the parameters of the log-likelihood ratio calibration module, so that the log-likelihood ratio calibration module can participate in joint training and perform end-to-end fine-tuning on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module, so as to achieve joint optimization and adaptation of modulation and coding. When the joint loss function reaches the preset convergence condition or the training rounds reach the preset threshold, the trained modulator and demodulator parameters are saved.

5. The method as described in claim 1, characterized in that, The trained deep learning modulator is deployed at the transmitting end, and the encoded bitstream to be transmitted is geometrically constellated shaped and mapped before transmission, including: The transmitting device loads the trained modulator parameters and deploys the network weights of the deep learning modulator obtained from offline training to the baseband processing module of the transmitter. Channel coding is performed on the user data to be transmitted, and the encoded bit stream is divided into bit groups according to the modulation order; The divided bit groups are input into the deep learning modulator, which maps each bit group into a constellation symbol after geometric constellation shaping according to the trained parameters. The constellation symbol presents a non-uniformly distributed shape on the complex plane, and the distance between different constellation points is adaptively adjusted according to the protection requirements of the bits they carry. The generated constellation symbols are transmitted to the receiving end via a wireless channel.

6. The method as described in claim 1, characterized in that, The trained deep learning demodulator is deployed at the receiver to demodulate the received signal. The calibrated log-likelihood ratio sequence is output by the log-likelihood ratio calibration module to the channel decoder for decoding, including: The receiving device loads the trained demodulator parameters and deploys the network weights of the deep learning demodulator obtained from offline training to the signal processing module of the receiver. The network weights include the trainable scaling parameters and bias parameters of the log-likelihood ratio calibration module. The received signal is subjected to channel estimation and equalization processing. The channel state information is estimated using the pilot signal, and the equalization algorithm is used to eliminate the phase rotation and amplitude distortion caused by channel fading, so as to obtain the equalized signal. The equalized signal is input into the deep learning demodulator. After feature extraction by the basic demodulation module, the log-likelihood ratio calibration module performs affine transformation on the extracted raw log-likelihood ratio and directly outputs the log-likelihood ratio sequence for each encoded bit. The log-likelihood ratio sequence is input into the channel decoder, which uses the belief propagation algorithm to perform iterative decoding to recover the original user information bits.

7. A deep learning-based geometric constellation shaping and joint coding modulation device, characterized in that, include: The joint coding and modulation system construction module is used to obtain the initial communication system model, which is configured to include a deep learning modulator at the transmitting end, a channel model, and a deep learning demodulator at the receiving end. The deep learning demodulator includes a basic demodulation module and a log-likelihood ratio calibration module integrated at the end. The log-likelihood ratio calibration module is configured to calibrate the raw log-probability output of the basic demodulation module to a log-likelihood ratio sequence that matches the input distribution of the channel decoder, and outputs the completed end-to-end communication system model. The multi-objective joint loss function construction module is used to obtain the joint loss function, configure the joint loss function to include a cross-entropy loss term for measuring bit error rate, a generalized mutual information loss term for maximizing channel capacity, and a boundary constraint loss term for constraining log-likelihood ratio confidence, and output the determined multi-objective joint loss function. The phased joint training strategy execution module is used to execute a phased joint training strategy using the end-to-end communication system model and the multi-objective joint loss function. In the first phase, the parameters of the log-likelihood ratio calibration module are frozen, and only the weights of the deep learning modulator and the basic demodulation module are updated to perform basic constellation shaping. In the second phase, the parameter freezing state of the log-likelihood ratio calibration module is lifted, and end-to-end joint fine-tuning is performed on the entire communication link including the deep learning modulator, the basic demodulation module and the log-likelihood ratio calibration module until the joint loss function converges, and the trained deep learning modulator parameters and demodulator parameters are output. The deep learning modulator deployment and mapping transmission module is used to deploy the trained deep learning modulator parameters at the transmitting end, use the deep learning modulator to perform geometric constellation shaping mapping on the encoded bit stream to be transmitted, and send the mapped constellation symbols to the receiving end via a wireless channel. The deep learning demodulator deployment and LLR decoding module is used to deploy the trained deep learning demodulator parameters at the receiver, demodulate the received signal using the deep learning demodulator, output the calibrated log-likelihood ratio sequence from the log-likelihood ratio calibration module, and output the log-likelihood ratio sequence to the channel decoder for decoding, outputting the recovered user information bits.

8. An electronic device, comprising: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the deep learning-based geometric constellation shaping joint coding modulation method as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements a deep learning-based geometric constellation shaping joint coding modulation method as claimed in any one of claims 1-6.

Images