Deep learning communication system optimization method based on channel perception

Abstract

The invention discloses a deep learning communication system optimization method based on channel perception. The invention provides a deep learning communication system optimization method based on channel perception, which introduces instantaneous channel state information of real physical channel characteristics in a model training process, and realizes physical layer driving optimization of a deep learning communication model through a channel perception mechanism. The method comprises the following steps: acquiring an instantaneous frequency response of a real or simulated channel, calculating an effective signal-to-noise ratio and a corresponding bit error rate of each subcarrier, and generating a non-uniform bit error rate vector; according to the vector, performing structured error injection on a bit stream in a training stage, constructing a dual-path loss function containing a noise path and a clean path, and performing end-to-end joint optimization on an encoder, a quantizer, a dequantizer and a decoder of a communication model; and through double-path weighted learning, the channel robustness of the model is enhanced while the training stability is ensured.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to an optimization method for a deep learning communication system, particularly a channel-aware model training method. Background Technology

[0002] With the advent of the era of "Internet of Everything," deep learning technology is profoundly reshaping the design paradigm of modern wireless communication systems. Traditional communication systems follow the design principle of separating source and channel coding. Although theoretically optimal, in practical applications, especially in environments with variable or harsh channel conditions, their performance often drops sharply due to the "cliff effect." To overcome this bottleneck, end-to-end communication architectures based on deep learning have emerged.

[0003] End-to-end schemes, exemplified by Joint Source-Channel Coding (JSCC), integrate and jointly optimize modules such as source coding, channel coding, and even modulation through neural networks. This enables smoother and more elegant performance transitions when channel conditions change, demonstrating enormous application potential. To ensure the compatibility of these advanced end-to-end systems with the currently widely deployed bitstream-based digital communication infrastructure, researchers have introduced quantization techniques (especially vector quantization) to successfully transform the continuous features extracted by the encoder into discrete bitstreams for transmission, achieving so-called "digital semantic communication."

[0004] However, despite the aforementioned progress in digital end-to-end communication systems, their development still faces a core and pervasive technical bottleneck: the disconnect between the training environment and the real physical environment. Existing research, in designing and training deep learning communication models, largely relies on oversimplified channel models, such as additive white Gaussian noise channels, Rayleigh or Ricean flat fading channels. Even some works attempt to introduce trainable bit-flipping probabilities to simulate digital channels, but these essentially equate the physical channel to a parallel binary symmetric channel; such models fail to reflect the inherent correlations of real channels.

[0005] These simplified models all fail to accurately capture the frequency-selective fading characteristics prevalent in real physical channels caused by multipath propagation effects. More importantly, they neglect structured, bursty bit error patterns closely related to instantaneous channel state information. This huge gap between "model idealization" and "real-world complexity" leads to the so-called "simulation-to-reality" performance gap. Communication systems trained on simplified channel models, once deployed in real physical environments where channel characteristics vary drastically with frequency and time, exhibit significantly reduced robustness and transmission efficiency, failing to fully realize their theoretical performance. Summary of the Invention

[0006] The purpose of this invention is to provide a channel-aware deep learning communication system optimization method, which aims to solve the technical problem that existing deep learning communication models cannot effectively combat frequency-selective fading and structured errors in real physical channels due to their reliance on simplified channel models, resulting in insufficient robustness of data transmission.

[0007] According to one aspect of the present invention, a channel-aware deep learning communication system optimization method is provided, applied to training a deep learning communication system model including an encoder, a quantizer, a dequantizer, and a decoder, comprising:

[0008] S1. Obtain instantaneous channel state information. In each iteration of model training, an instantaneous channel frequency response vector is randomly sampled from a channel database that reflects the real physical channel characteristics. ,in This vector represents the total number of subcarriers. It contains channel gain information for multiple subcarriers.

[0009] S2. Generate structured error injection parameters. Based on the instantaneous channel frequency response vector obtained in step S1. and the average signal-to-noise ratio of the current training The effective signal-to-noise ratio and corresponding bit error rate of each subcarrier are calculated to generate a non-uniform bit error rate vector that directly reflects the frequency selectivity structure of the channel. .

[0010] S3. Construct a dual-path loss function and train the model. The source data, after passing through an encoder (i.e., a deep learning network) and quantization, yields a raw, error-free bitstream. It is then processed through the following two paths:

[0011] Noise path: converting the original bitstream Input a structured error injection module, which performs bit flipping on the original bitstream based on the bit error rate vector generated in step S2 to generate an error-injected bitstream. The mis-transmitted bit stream The input dequantizer and decoder (also deep learning networks, representing the inverse operation of the encoder) are used to reconstruct the data, and the noisy path loss used to evaluate the model's robustness is calculated. .

[0012] Clean path: the raw bitstream that has not been injected with errors. The dequantizer and decoder are directly input, and the clean path loss used to stabilize the training process and optimize the quantizer infrastructure is calculated. .

[0013] S4. Update model parameters. (Include noisy path loss) With clean path loss We perform a weighted summation to obtain the total loss function. The gradient based on the total loss function The parameters of the encoder, decoder, quantizer and dequantizer in the communication model are jointly optimized by using the backpropagation algorithm.

[0014] Furthermore, in step S1, the instantaneous channel frequency response vector Channel gain amplitude of each subcarrier The significant differences due to frequency-selective fading directly reveal which subcarriers are in a state of deep fading.

[0015] Further, in step S2, the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier are calculated, including: first, according to the formula Calculate the first Effective signal-to-noise ratio of each subcarrier Then, the effective signal-to-noise ratio is input into a preset theoretical bit error rate calculation function to obtain the bit error probability of the subcarrier. Although different modulation schemes correspond to different bit error rate (BER) calculation functions, the error probability here only reflects one aspect of channel error conditions. Therefore, the impact of the choice of BER calculation function on the overall system training results can be temporarily disregarded. Repeatedly calculating the bit error probabilities of all subcarriers yields the BER vector. The variation of this non-uniform BER vector directly encodes the frequency-selective fading mode of the channel.

[0016] Furthermore, in step S3, the noise path loss The expression is as follows: , It consists of two parts: It is the reconstruction loss that measures the difference between the original data and the reconstructed data. and These represent the source data and the data obtained after reconstruction by the decoder, respectively. The VQ loss measures the distance between the original feature vector and the erroneous feature vector caused by bit errors. and These represent the feature vector obtained by encoding the input data through the encoder and the erroneous feature vector input to the decoder after quantization, channel transmission, and dequantization operations, respectively. This composite loss enhances the robustness of the system by training it with corrupted inputs.

[0017] The reconstruction loss Used to measure raw data Compared with the reconstructed data obtained after decoding the noisy path The loss function is determined by the degree of distortion between the data points. It can be flexibly constructed from one or more distortion metrics chosen based on the specific data type and application requirements. Its core purpose is to guide the decoder to learn how to recover the original data from features affected by channel impairments with the highest possible quality.

[0018] The formula for calculating the VQ loss is: in This indicates that the gradient operation is stopped. Ensure It is still close to its target vector in the feature space. This enhances the structural robustness of the alignment flip. The encoder is updated to generate a representation that is less sensitive to channel-induced disturbances.

[0019] Furthermore, in step S3, the clean path loss It consists only of the standard VQ loss, and its expression is: in This represents a feature vector input to the decoder that is only affected by quantization and dequantization operations, and not by channel errors. Because its calculation excludes random channel errors, its gradient variance is low, playing a role in gradient variance suppression during training, thus providing a stable and noise-free gradient signal for the entire training process.

[0020] Furthermore, in step S4, the hyperparameters This is used to balance the stability and robustness of model training. All optimizable parameters are: The formula is as follows: , Find the optimal parameters .

[0021] According to one aspect of the present invention, a channel-aware deep learning communication system optimization device is provided, applied to training a deep learning communication system model including an encoder, a quantizer, a dequantizer, and a decoder, comprising:

[0022] The first main module is used to obtain the instantaneous channel frequency response vector of the real or simulated channel;

[0023] The second main module is used to calculate the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier in the channel based on the instantaneous channel frequency response vector, and generate a non-uniform bit error rate vector.

[0024] The third main module is used to feed the raw bitstream obtained after the encoder and quantizer process the source data into the noise path and the clean path in parallel. The noise path performs bit flipping on the raw bitstream according to the bit error rate vector to inject structured errors, generating an error-injected bitstream. This error-injected bitstream is then input into the dequantizer and decoder for data reconstruction, and the noise path loss is calculated accordingly. The clean path directly inputs the raw bitstream without error injection into the dequantizer and decoder, and the clean path loss is calculated accordingly.

[0025] The fourth main module is used to update the model parameters, perform a weighted summation of the noisy path loss and the clean path loss to obtain the total loss function, and perform joint optimization of the parameters in the deep learning communication system model based on the gradient of the total loss function.

[0026] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned channel-aware deep learning communication system optimization method.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. Significantly improves robustness in real channel environments: This invention introduces a physical layer structured error injection module driven by instantaneous channel state information during the training process, enabling deep learning communication models to learn feature representations that are highly resistant to frequency-selective fading and burst errors.

[0029] 2. Enhanced model training stability and convergence: The dual-path learning strategy designed in this invention effectively suppresses gradient variance by weighting and combining the gradients of noisy paths with high variance with those of noiseless paths with low variance. This makes the entire model training process more stable, converges faster, and ultimately yields a communication model that achieves a better balance between compression efficiency and channel robustness.

[0030] 3. Achieves high compatibility with existing digital communication systems: By representing information as a bit stream, it can be seamlessly integrated with existing digital modem modules. Through end-to-end joint optimization, compared to traditional methods, the model trained by this method can achieve the same or even better data recovery quality with lower transmission overhead (i.e., fewer bits). Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating a channel-aware deep learning communication system optimization method provided in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the overall framework of a communication system applied in a preferred embodiment of the present invention, showing the main functional modules including a transmitter, a wireless channel, and a receiver.

[0034] Figure 3 This is a schematic diagram of the core process for model training in a preferred embodiment of the present invention, which is used to illustrate a specific implementation process of the channel-aware optimization method proposed in the present invention.

[0035] Figure 4 This is a flowchart illustrating the deployment and testing of a trained model in a preferred embodiment of the present invention, demonstrating the signal transmission process in a real physical channel.

[0036] Figure 5 This is a performance comparison curve of a preferred embodiment of the present invention (specifically an image transmission application) and several comparative schemes under different signal-to-noise ratios, used to verify the beneficial effects of the present invention. Detailed Implementation

[0037] Existing deep learning communication models, relying on simplified channel models, are unable to effectively combat frequency-selective fading and structured errors in real physical channels, resulting in insufficient robustness in data transmission. Therefore, there is an urgent need in this field for a novel optimization method for deep learning communication systems that can bridge this gap. This method should enable the communication model to "sense" the complex characteristics of real physical channels during the training phase and learn robust strategies sufficient to withstand structured and sudden channel impairments, thereby significantly improving the reliability and performance of next-generation intelligent communication systems in practical applications.

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0039] This invention provides a channel-aware deep learning communication system optimization method, applied to training a deep learning communication system model including an encoder, quantizer, dequantizer, and decoder. Please refer to [link to relevant documentation]. Figure 1 The process includes: acquiring the instantaneous channel frequency response vector of a real or simulated channel; calculating the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier in the channel based on the instantaneous channel frequency response vector to generate a non-uniform bit error rate vector; feeding the raw bit stream obtained after the encoder and quantizer process the source data into a noise path and a clean path in parallel, wherein the noise path performs bit flipping on the raw bit stream according to the bit error rate vector to inject structured errors, generating an error-injected bit stream, and inputting the error-injected bit stream into a dequantizer and a decoder for data reconstruction, thereby calculating the noise path loss; the clean path directly inputs the raw bit stream without error injection into the dequantizer and decoder, thereby calculating the clean path loss; updating the model parameters, weighting and summing the noise path loss and the clean path loss to obtain a total loss function, and jointly optimizing the parameters in the deep learning communication system model based on the gradient of the total loss function.

[0040] To enable those skilled in the art to more clearly understand the technical solution of this invention, a deep learning communication system applied to joint source-channel coding (JSCC) for images will be described in detail below as a preferred embodiment of this invention. It should be emphasized that this embodiment is only used to illustrate the technical concept of this invention and does not constitute any limitation on the scope of protection of this invention. The core optimization method proposed in this invention is also applicable to deep learning communication systems that process audio, video, text, and other types of data.

[0041] The specific implementation of this invention includes an introduction to the overall system architecture, as well as two core parts: model training and model deployment. Please refer to [link / reference]. Figure 2The technical solution of this invention is applied to a digital JSCC communication system comprising a transmitter, a wireless channel, and a receiver. This system architecture serves as the foundational platform for the training method and testing verification of this invention.

[0042] The transmitter is responsible for transmitting an input source image Processed into OFDM complex symbol sequences suitable for transmission over digital wireless channels. The process includes: 1. Semantic encoder: Image First, it is fed into a deep neural network encoder. , For the parameters of the neural network to be trained, the encoder extracts key semantic features of the image and compresses them into a compact, continuous-valued latent vector. This process can be represented by the following expression: .

[0043] 2. Quantization Module: For compatibility with digital communication systems, continuous latent vectors... It needs to be quantized into a discrete bit stream. This embodiment uses vector quantization to... Divide into N sub-vectors and in the trainable codebook Find the nearest codeword index for each subvector. Use the following formula: , This represents the most recently found codeword. After performing the above operations on all subvectors, an index set that can be losslessly mapped to a binary bitstream can be generated. .

[0044] 3. Modulation module: In actual deployment, the index set The data will be losslessly converted into a bitstream, which will then be modulated into OFDM complex symbols by a modulation module. It is then transmitted to the wireless channel. No modulation module is used during training.

[0045] The wireless channel is used to establish a communication link between the transmitter and the receiver. It should be noted that, in the implementation of this invention, the simulation and processing methods of the wireless channel differ for the two different stages of model training and model deployment; specific details will be described later.

[0046] Receiver: The receiver is responsible for recovering the source image as accurately as possible from the received, possibly damaged, signal. This process includes: 1. Demodulation module: In actual deployment, it demodulates the received signal. Demodulation is performed to obtain the estimated codeword index. .

[0047] 2. Dequantization module: using indexes In the codebook shared with the transmitter The corresponding codeword is retrieved from the database, thereby reconstructing the estimated value of the latent vector. .

[0048] 3. Semantic decoder: reconstructs the latent vectors Input to semantic decoder , Using the parameters of the neural network to be trained, the reconstructed image is recovered. This process can be represented by the following expression: .

[0049] The core innovation of this invention lies in its unique training phase, which does not involve real channel transmission. Instead, it uses the physical layer structured error injection module and dual-path loss function designed in this invention to perform efficient and realistic optimization of the encoder, decoder, quantizer and dequantizer of the above system.

[0050] Please see Figure 3 This section details the complete training process of the communication model used in the embodiments of the present invention.

[0051] In this embodiment, the source data used for training is the CIFAR-10 image dataset, which contains 60,000 color images with a size of 32×32 pixels. Preferably, both the encoder and decoder of the communication model adopt a deep residual network structure and integrate a self-attention module to efficiently capture long-range dependencies in the feature maps. Specifically, the encoder compresses each 32×32×3 image into N=256 latent subvectors with a dimension of D=64. Of course, the present invention is not limited to this; other neural network structures with feature extraction and reconstruction functions, such as Transformer, are also applicable.

[0052] To train the model to adapt to real-world channel environments, a channel database reflecting the physical channel characteristics is needed. In this embodiment, to facilitate large-scale, repeatable experimental verification, the channel database is generated through simulation using QuaDriGa software conforming to the 3GPP 38.901 standard, containing 1000 independent sets of OFDM channel data, each containing 48 subcarriers. However, those skilled in the art should understand that this is merely an example, and the method proposed in this invention is not limited to simulated channel data. Its core lies in using instantaneous channel state information to drive error injection; therefore, this method is also applicable to channel state information data collected in real-world environments using physical channel measurement equipment.

[0053] The model training process strictly follows the method proposed in this invention. During training, a multi-codebook adaptive strategy is employed to enhance the model's adaptability to different channel qualities. Specifically, five codebooks are set up, trained for five different signal-to-noise ratio (SNR) levels (20dB, 15dB, 10dB, 5dB, 0dB), with only one codebook activated during each training iteration. Except for the first codebook (20dB), which is randomly initialized, each subsequent codebook is initialized by copying the previously trained codebook to accelerate convergence. The codeword index bits for each codebook are set to W=9, and training lasts for 100 epochs. The Adam optimizer is used for training, with an initial learning rate set to... The system dynamically adjusts the hyperparameters based on the signal-to-noise ratio during the current training phase. In this embodiment, the balancing hyperparameters in the dual-path loss function... Set to 0.7, the weight hyperparameter in the reconstruction loss. Set it to 0.1.

[0054] The model is trained in an iterative loop, and each iteration strictly follows the core method proposed in this invention. The specific process is as follows: Feature encoding and quantization: Encoding an input image Input by parameters Semantic encoder of control To obtain the latent vector of continuous values Subsequently, a vector quantization method was used to... Divided into 256 sub-vectors For each subvector Through the currently active codebook Find the codeword that minimizes the Euclidean distance and obtain its index. All the indices together form a raw index set that can be losslessly mapped to a binary bit stream. .

[0055] Parallel dual-path processing: index set The data is losslessly mapped to a bitstream and then enters a parallel dual-path processing stage, which corresponds to the process of the bitstream passing through the wireless channel during training.

[0056] (1) Noise path: Step S1: Obtain instantaneous channel state information by randomly sampling an instantaneous channel frequency response vector from the channel database generated above. .

[0057] Step S2, generate structured error injection parameters, based on this vector and the average signal-to-noise ratio of the current training Calculate the effective signal-to-noise ratio for each subcarrier. Preferably, to obtain a simple and easily computed channel quality mapping, this embodiment substitutes the theoretical bit error rate function of BPSK. The bit error probability of the subcarrier is obtained. , here It is the tail probability function of the standard normal distribution. This generates a non-uniform bit error rate vector that directly reflects the channel's frequency selectivity structure. .

[0058] Step S3: Logically map the bitstream onto each subcarrier, and based on the bit error rate vector mentioned above, probabilistically flip each bit using independent Bernoulli trials to generate an index sequence containing structured errors. The index sequence is dequantized to obtain a feature vector containing errors. .

[0059] (2) Clean path: Step S3, the original index set that has not been injected with any errors. The codewords are directly fed into the dequantization module to obtain the reconstructed codewords. .

[0060] Loss calculation and parameter update: Step S3: Calculate the loss for each of the two paths. (Noisy path loss) Reconstruction loss and VQ loss Composition; Clean path loss Standard VQ loss only constitute.

[0061] Preferably, the reconstruction loss It is a weighted sum of mean squared error (MSE) and perceived loss (LPIPS), expressed as: in It is a hyperparameter used to balance the weights of the two parts. Ensure basic pixel-level fidelity, while Enhancing visual realism. Both work together to ensure the decoder can reconstruct high-quality images from error-prone features.

[0062] Preferably, the VQ loss The calculation formula is: in This indicates that the gradient operation is stopped. Update the codebook to ensure It is still close to its target vector in the feature space. This enhances the structural robustness of the alignment flip. The encoder is updated to generate a representation that is less sensitive to channel-induced disturbances.

[0063] Preferably, the standard VQ loss The expression is: in This represents the codeword retrieved from the codebook using an ideal and undamaged index. Because its calculation excludes random channel errors, its gradient variance is low, playing a role in gradient variance suppression during training, thus providing a stable and noise-free gradient signal for the entire training process.

[0064] Step S4, based on the total loss function Calculate its parameters with respect to all parameters of the model. The gradient is calculated, and the encoder parameters are processed using the backpropagation algorithm. Decoder parameters and codebook parameters Perform joint optimization. Preferably, all optimizable parameters are: The formula is as follows: , Find the optimal parameters .

[0065] Please see Figure 4 Preferably, this section describes the deployment of the model trained in the first part, and objectively and comprehensively verifies the superiority of the model trained in this embodiment by comparing it with a variety of existing technologies.

[0066] In actual deployment, the parallel dual-path processing method described in the training phase will not be used; the OFDM complex symbol sequence obtained by the modulation module in the transmitter will be... The transmission will take place through an actual physical channel, and the expression for this process is as follows: in This is the channel frequency response vector. Given an additive white Gaussian noise vector, we obtain... The input will be sent to the demodulation module of the receiver, which corresponds to the complete process of the wireless channel during the deployment phase. Meanwhile, in this embodiment, to accurately simulate physical layer transmission, the transmitter's modulation module employs an adaptive resource scheduling strategy. This strategy is based on the rate maximization principle, independently allocating the optimal modulation order and transmit power to each OFDM subcarrier according to the instantaneous CSI and total power budget.

[0067] To fully verify the beneficial effects of the present invention, the following representative comparative schemes are set up in this embodiment: AWGN-JSCC: This represents a mainstream approach in the prior art. Its network structure is consistent with that of this embodiment, but it is trained on a simplified additive white Gaussian noise channel model.

[0068] SQ-JSCC: The JSCC scheme replaces the vector quantization scheme in this embodiment with the scalar quantization scheme.

[0069] JPEG-10 / 20+LDPC: This represents a traditional technique that separates source and channel coding, where 10 and 20 indicate different JPEG compression rates.

[0070] Please see Figure 5 The figure illustrates a comparison of PSNR performance metrics between a specific embodiment of the method of this invention (labeled "PDP-JSCC") and the aforementioned comparative scheme within the 0dB to 20dB signal-to-noise ratio range. The figure shows that the PDP-JSCC scheme significantly outperforms both the AWGN-JSCC scheme trained on a simplified AWGN channel and the SQ-JSCC scheme using scalar quantization. Compared to the JPEG+LDPC scheme, the PDP-JSCC scheme exhibits a significant performance advantage in the low signal-to-noise ratio region, avoiding the "cliff effect" and achieving smooth performance degradation. This fully demonstrates that the optimization method proposed in this invention, by introducing a structured error injection module that simulates the characteristics of real physical channels for training, enables deep learning communication models to achieve stronger robustness, thereby achieving better data recovery quality in practical applications. Specifically, in the image transmission task of this embodiment, this manifests as a significant improvement in image reconstruction quality.

[0071] The implementation of the various embodiments of the present invention is based on programmed processing through a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the present invention provides a channel-aware deep learning communication system optimization device, which is used to execute a channel-aware deep learning communication system optimization method from the above method embodiments.

[0072] This device is used to train a deep learning communication system model that includes an encoder, quantizer, dequantizer, and decoder. It includes: a first main module for acquiring the instantaneous channel frequency response vector of a real or simulated channel; a second main module for calculating the effective signal-to-noise ratio (SNR) and corresponding bit error rate (BER) of each subcarrier in the channel based on the instantaneous channel frequency response vector, generating a non-uniform BER vector; a third main module for feeding the raw bitstream obtained after processing the source data by the encoder and quantizer into a noise path and a clean path in parallel. The noise path performs bit flipping on the raw bitstream according to the BER vector to inject structured errors, generating an error-injected bitstream, which is then input into the dequantizer and decoder for data reconstruction, thereby calculating the noise path loss; the clean path directly inputs the raw bitstream without error injection into the dequantizer and decoder, thereby calculating the clean path loss; and a fourth main module for updating model parameters by weighted summing of the noise path loss and the clean path loss to obtain a total loss function, and jointly optimizing the parameters in the deep learning communication system model based on the gradient of the total loss function.

[0073] This invention provides a channel-aware deep learning communication system optimization device. Addressing the technical problem that existing deep learning communication models, relying on simplified channel models, cannot effectively combat frequency-selective fading and structured errors in real physical channels, resulting in insufficient data transmission robustness, this invention employs the aforementioned modules to enable the communication model to "aware" the complex characteristics of real physical channels during the training phase and learn robust strategies sufficient to resist structured and sudden channel impairments. This significantly improves the reliability and performance of next-generation intelligent communication systems in practical application scenarios.

[0074] It should be noted that the device embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above device embodiments provided by the present invention. As long as those skilled in the art can improve the device in the above device embodiments by referring to the specific technical solutions in other method embodiments, combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, and ensuring the practicality of the technical solutions, they can obtain corresponding device-type embodiments for implementing the methods in other method-type embodiments.

[0075] Based on the same inventive concept as any of the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned channel-aware deep learning communication system optimization method. The method is applied to training a deep learning communication system model including an encoder, a quantizer, a dequantizer, and a decoder.

[0076] In summary, this invention provides a channel-aware deep learning communication system optimization method. It introduces instantaneous channel state information based on real physical channel characteristics during model training, and achieves physical layer-driven optimization of the deep learning communication model through a channel-aware mechanism. The method includes: acquiring the instantaneous frequency response of a real or simulated channel; calculating the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier; generating a non-uniform bit error rate vector; injecting structured errors into the bitstream during the training phase based on this vector; constructing a dual-path loss function containing noisy and clean paths; and performing end-to-end joint optimization of the encoder, quantizer, dequantizer, and decoder of the communication model; and enhancing the model's channel robustness through dual-path weighted learning while ensuring training stability. This method allows the model to fully "aware" the frequency-selective fading and burst error characteristics of the real channel during the training phase, thereby significantly improving transmission reliability and anti-interference capability in complex physical channel environments, and achieving efficient and stable operation of the deep learning communication system in real-world deployments.

[0077] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A channel-aware deep learning communication system optimization method, applied to training a deep learning communication system model including an encoder, quantizer, dequantizer, and decoder, characterized in that, include: Obtain the instantaneous channel frequency response vector of a real or simulated channel; Based on the instantaneous channel frequency response vector, the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier in the channel are calculated to generate a non-uniform bit error rate vector. The raw bitstream obtained after the encoder and quantizer process the source data is fed into the noise path and the clean path in parallel. The noise path performs bit flipping on the raw bitstream according to the bit error rate vector to inject structured errors, generating an error-injected bitstream. This error-injected bitstream is then input into the dequantizer and decoder for data reconstruction, and the noise path loss is calculated accordingly. The clean path directly inputs the raw bitstream without error injection into the dequantizer and decoder, and the clean path loss is calculated accordingly. The model parameters are updated by weighted summing of the noisy path loss and the clean path loss to obtain the total loss function, and the parameters in the deep learning communication system model are jointly optimized based on the gradient of the total loss function.

2. The channel-aware deep learning communication system optimization method according to claim 1, characterized in that, Calculate the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier in the channel, including: Based on the average signal-to-noise ratio and the instantaneous channel frequency response vector Channel gain corresponding to subcarrier Calculate the effective signal-to-noise ratio of the subcarrier. ; The effective signal-to-noise ratio Input the preset theoretical bit error rate calculation function to obtain the bit error rate of the subcarrier.

3. The channel-aware deep learning communication system optimization method according to claim 1, characterized in that, The noise path loss is calculated by comparing the original data x with the reconstructed data. Reconstruction loss of differences and measures the original feature vector Error feature vectors caused by bit errors VQ loss of the distance between Together they constitute.

4. The channel-aware deep learning communication system optimization method according to claim 3, characterized in that, The expression for the noise path loss is as follows: 。 5. The channel-aware deep learning communication system optimization method according to claim 3, characterized in that, The clean path loss consists solely of VQ loss and is used to provide a stable, low-variance gradient signal during training.

6. The channel-aware deep learning communication system optimization method according to claim 5, characterized in that, The expression for the clean path loss is as follows: ， in This indicates that the gradient operation is stopped. This represents a feature vector input to the decoder that is only affected by quantization and dequantization operations, and not by channel errors. and The difference comes from the quantization and dequantization operations.

7. The channel-aware deep learning communication system optimization method according to claim 5, characterized in that, The expression for the VQ loss is as follows: ， in This indicates that the gradient operation is stopped. make sure It is still close to its target vector in the feature space. , The encoder is updated to generate a representation that is less sensitive to channel-induced disturbances.

8. The channel-aware deep learning communication system optimization method according to claim 7, characterized in that, The expression for the total loss function is as follows: ， in, This is a hyperparameter.

9. A channel-aware deep learning communication system optimization device, used to train a deep learning communication system model including an encoder, quantizer, dequantizer, and decoder, characterized in that, include: The first main module is used to obtain the instantaneous channel frequency response vector of the real or simulated channel; The second main module is used to calculate the effective signal-to-noise ratio and corresponding bit error rate of each subcarrier in the channel based on the instantaneous channel frequency response vector, and generate a non-uniform bit error rate vector; The third main module is used to feed the raw bitstream obtained after the encoder and quantizer process the source data into the noise path and the clean path in parallel. The noise path performs bit flipping on the raw bitstream according to the bit error rate vector to inject structured errors, generating an error-injected bitstream. This error-injected bitstream is then input into the dequantizer and decoder for data reconstruction, and the noise path loss is calculated accordingly. The clean path directly inputs the raw bitstream without error injection into the dequantizer and decoder, and the clean path loss is calculated accordingly. The fourth main module is used to update the model parameters, perform a weighted summation of the noisy path loss and the clean path loss to obtain the total loss function, and perform joint optimization of the parameters in the deep learning communication system model based on the gradient of the total loss function.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the channel-aware deep learning communication system optimization method according to any one of claims 1 to 8.

Images