Digital semantic communication method combining learnable constellation mapping and channel adaptation

Abstract

The invention relates to a digital semantic communication method combining learnable constellation mapping and channel adaptation, and belongs to the technical field of communication. According to the method, a semantic coding modulator is constructed, and the transition probability from an information source to a discrete constellation symbol is systematically learned, so that the problem that digital modulation cannot be differentiated is solved; a semantic coding process is dynamically adjusted based on a channel state information feedback mechanism of double attention, and an efficient matching relationship between semantic features and subcarriers is ensured. A neural network is designed to combine a multipath channel which cannot be trained but can be differentiated with an OFDM signal processing module, and end-to-end cross-layer joint optimization of the system is realized.

Description

Technical Field

[0001] This invention belongs to the field of communication technology and relates to a digital semantic communication method that combines learnable constellation mapping and channel adaptation. Background Technology

[0002] Semantic communication, as a new type of communication mode, is based on the idea of ​​extracting and transmitting task-related semantic information and eliminating redundant and invalid information, thereby effectively reducing the data transmission burden and improving communication efficiency. It has important application value in future communication networks.

[0003] In recent years, research by scholars both domestically and internationally on the architecture and application scenario design of semantic communication systems has mainly focused on three aspects: analog semantic communication, digital semantic communication, and hybrid semantic communication. In analog semantic communication, end-to-end designs based on neural network-based semantic codecs are employed. These encoders typically output continuously distributed signals, which are then directly fed into the communication channel. These signals are not modulated into discrete constellation symbols, making it difficult to deploy semantic communication systems on modern digital communication equipment. Therefore, considering the complexity of hardware implementation and compatibility with existing digital communication protocols, designing digital semantic communication systems has become crucial for a smooth transition from traditional communication paradigms to semantic communication paradigms.

[0004] In deep learning-based end-to-end communication systems, establishing a usable channel model is crucial for the learning process of optimization algorithms. However, when the channel characteristics are unknown a priori, the system faces the optimization dilemma of backpropagation gradients being interrupted due to the non-differentiability of the channel. Therefore, existing research mainly focuses on simple channel models, such as additive white Gaussian noise channels and Rayleigh flat fading channels. This approach often ignores complex channel models, such as multipath channels, which introduce challenges like inter-symbol interference (ISI), significantly reducing semantic communication performance.

[0005] In recent years, significant efforts have been made to combine end-to-end learning-based communication systems with Orthogonal Frequency Division Multiplexing (OFDM) as domain knowledge. However, previous research has focused only on the design of modulation constellations for learning arbitrary binary sequence inputs, without considering neural network-based source coding.

[0006] Furthermore, current deep learning-driven Joint Source-Channel Coding (JSCC) methods still face significant challenges in terms of channel adaptability. Existing schemes mainly employ two training paradigms: one is a dedicated model for a fixed signal-to-noise ratio (SNR), which requires storing a large number of channel condition-dependent network parameters; the other is training for a certain range of SNRs, which saves storage space but leads to performance degradation and fails to effectively utilize Channel State Information (CSI). In contrast, traditional digital communication systems have demonstrated that CSI can not only optimize communication rates through power allocation at the transmitter, but also guide intelligent feature-subcarrier mapping in OFDM systems, such as prioritizing the allocation of key semantic features of images to high-reliability subcarriers.

[0007] Therefore, existing digital semantic communication technologies still face pressing technical challenges in addressing the non-differentiability of digital modulation, effectively adapting to time-varying channels using limited transmission resources, and overcoming signal distortion caused by multipath fading. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a digital semantic communication method that combines learnable constellation mapping and channel adaptation. By constructing a semantic encoder-modulator, the system learns the transition probabilities from the source to discrete constellation symbols, thereby solving the problem of non-differentiable digital modulation. A dual-attention channel state information (CSI) feedback mechanism dynamically adjusts the semantic coding process, ensuring an efficient matching relationship between semantic features and subcarriers. A neural network is designed to combine an untrainable but differentiable multipath channel with an orthogonal frequency division multiplexing (OFDM) signal processing module, achieving end-to-end cross-layer joint optimization of the system.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A digital semantic communication method that combines learnable constellation mapping and channel adaptation includes the following steps: S1: Input the source data into the semantic coding modulator to obtain a differentiable constellation symbol; S2: Apply average power constraints to the signal using a power constraint unit; S3: The signal is fed into the OFDM transmitter to perform an inverse fast fourier transform (IFFT) and add a cyclic prefix (CP), which is then mapped to a baseband signal. S4: Passing through the wireless channel; S5: The signal is fed into the OFDM receiver for Discrete Fourier Transform (DFT) operation and CP removal, and channel estimation is performed. S6: Perform equalization using the channel estimation results, and feed the channel estimation results back to the semantic coding modulator and semantic decoder. S7: Reconstruct source data using a semantic decoder; S8: Calculate the loss function to jointly optimize the transition probabilities of the sender and receiver, and update the gradient.

[0010] Furthermore, in step S1, the semantic coding modulator includes a probabilistic encoder and a learnable constellation symbol mapping module for processing source data. Generate constellation symbol sequence The constellation symbol mapping module uses parameters... The neural network is parameterized to learn the transition probabilities:

[0011] in, This represents a semantic coding modulator consisting of a probabilistic encoder and a learnable constellation symbol mapping module. It is the source data, i.e., the input image. c , h and w These represent color, height, and width, respectively.

[0012] Furthermore, step S1 specifically includes: for The general transition probability model of the first-order modulation, taking a length of... Constellation symbol sequence , , each component Modeled as conditionally independent variables, where the components Choosing a constellation of One of the values ​​in a set of symbols, the neural network outputs each symbol. The classification distribution will It's a constellation symbol. The probability of one of them is expressed as:

[0013] Since it is assumed that the components are conditionally independent, the joint probability mass function of the semantic coding modulator is expressed as:

[0014] in, Indicates an indicator function, Indicates when yes The probability of one of the symbols in the diagram; Constellation symbols are generated using a reparameterization technique based on Gumbel-Softmax. End-to-end training is achieved through the following mechanism: during forward propagation, discrete constellation symbols are generated using Gumbel-Max sampling, and during backpropagation, gradient estimation is performed by replacing the argmax operation with a differentiable softmax approximation.

[0015] Furthermore, step S2 specifically includes: outputting the constellation symbol sequence. Reshaped into , This represents the number of subcarriers in an OFDM data symbol. This refers to the number of OFDM data symbols; the reshaped constellation symbol sequence is passed through a power constraint, and the signal is subject to the following average power constraint:

[0016] Among them, expectations It is calculated based on the input image. Represents the Frobenius norm; Represent the average power constraint; without loss of generality, assume .

[0017] Furthermore, step S3 specifically includes: converting OFDM symbols into OFDM signals using inverse fast Fourier transform (IFFT); and inserting CP into the OFDM signal to solve problems such as multipath channels and inter-symbol interference, thereby forming the transmitted frequency domain complex-valued signal. , is represented as:

[0018] in, It is the CP length. It represents the number of pilot symbols.

[0019] Furthermore, in step S4, after passing through the wireless channel, the signal received by the receiving end... Represented as:

[0020] in, Indicates the wireless channel. Indicates the convolution operation; This represents additive white Gaussian noise, where each path undergoes independent Rayleigh fading, satisfying... , , For the number of multipaths, Indicates the first l A path.

[0021] Furthermore, in step S5, the receiving end receives a noisy signal output from the channel. First, the cyclic prefix (CP) is removed, and then the data is transformed to the frequency domain using Discrete Fourier Transform (DFT) to obtain the pilot symbols. and data symbols ;Will and The following channel estimate is obtained by combining the results:

[0022] in, This represents the channel estimation result. Indicates the local pilot signal; It is the first k Channel estimation gain for each subcarrier .

[0023] Furthermore, in step S6, the data symbols after Discrete Fourier Transform (DFT) Using channel estimation results Achieve the following equilibrium:

[0024] The receiving end will equalize the data The data is fed into the semantic decoder to reconstruct the source data and complete the corresponding intelligent task; simultaneously, the estimated... The feedback will be fed into the dual attention module in the semantic encoder-modulator and semantic decoder. The dual attention module consists of two neural network architectures, each consisting of an average pooling layer, a fully connected layer, a PReLU layer, a fully connected layer, and a sigmoid layer. The purpose is to utilize CSI information for more adaptive power allocation.

[0025] Furthermore, step S7 specifically includes: receiving The semantic decoder estimates the posterior distribution. Then, the source data is recovered by decoding with maximum a posteriori probability: ,in This represents the parameters of the neural network.

[0026] Furthermore, in step S8, the optimization objective is to be able to receive the sequence To retain as much information as possible from the source data, we define an objective function based on maximizing mutual information, and formulate the optimization problem as follows:

[0027] Specifically, taking into account the optimization of the semantic decoder, a general loss function is designed to jointly optimize the transition probability at the sender. and receiver probability First, construct the true posterior distribution. An approximate inference model is derived, and then variational inference is used to derive the following lower bound based on the mutual information loss function:

[0028] This lower bound includes the true posterior distribution. With neural network parameterized posterior distribution The distance metric between them; where, Indicates that under given parameters Next, input image With the received equalized signal Statistical correlation between them; Indicates the received equalization signal Given a transmitted signal Y The expectation under the conditional distribution; Indicates the encoder-modulator output F Given an input image and parameters The expectation under the conditional distribution; Indicates the input image Given the received equalization signal The expectation under the posterior distribution; Indicates the input image The entropy.

[0029] The beneficial effects of this invention are as follows: (1) The present invention constructs a semantic coding modulator and uses a probabilistic modeling method to learn the transition probability from the source to the discrete constellation symbol, thereby effectively solving the non-differentiable problem of the traditional digital modulation process in the end-to-end deep learning system.

[0030] (2) By utilizing the CSI (Channel State Information) feedback mechanism based on dual attention, this method can dynamically capture the time and frequency domain characteristics of multipath channels and adjust the semantic coding process accordingly. This enables the system to prioritize the transmission of key semantic features on high-quality, high-reliability subcarriers, thereby improving the robustness and efficiency of communication.

[0031] (3) This invention ingeniously combines a neural network (semantic codec) with a non-trainable but differentiable multipath channel model and an OFDM (orthogonal frequency division multiplexing) signal processing module. This design enables end-to-end cross-layer joint optimization of the entire communication system (from source coding to channel transmission to signal reception), thereby overcoming the signal distortion problems caused by multipath fading and inter-symbol interference (ISI).

[0032] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the digital semantic communication method that combines learnable constellation mapping and channel adaptation according to the present invention; Figure 2 This is a flowchart of the system framework. Figure 3 A schematic diagram of a CSI feedback mechanism based on dual attention; Figure 4 This is a diagram of the neural network architecture for semantic encoding and decoding. Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0035] This invention provides a digital semantic communication method that combines learnable constellation mapping and channel adaptation, in conjunction with... Figures 1-4 The explanation includes the following steps: Step 1: Information Source. Taking images as an example, let's assume each image... In a containing pilot symbols and Transmitted in OFDM packets containing data symbols. At the sending end, source data carrying semantic information... Following the general objective of image semantic communication, the receiver needs to recover the original source data and represent the recovered source data as... The shared background knowledge base also includes source data. .

[0036] The semantic coding modulator consists of a probabilistic encoder and a learnable constellation symbol mapping module, used to extract information from source data. generate This module uses parameters. The neural network is parameterized to learn the transition probabilities:

[0037] in, It is an input image. c , h and w These represent color, height, and width, respectively.

[0038] The following section details the formula for the transition probability model and the generation of constellation symbol sequences.

[0039] for The general transition probability model of the first-order modulation, taking a length of... Constellation symbol sequence , , each component Modeled as conditionally independent variables, where the components Choosing a constellation of One of the values ​​in a set of symbols, the neural network outputs each symbol. The classification distribution will It's a constellation symbol. The probability of one of them is expressed as:

[0040] Since it is assumed that the components are conditionally independent, the joint probability mass function of the semantic coding modulator is expressed as:

[0041] in, Indicates an indicator function, Indicates when yes The probability of one of the symbols in the diagram.

[0042] This embodiment employs a Gumbel-Softmax-based reparameterization technique to generate constellation symbols. This method achieves end-to-end training through the following mechanism: during forward propagation, discrete constellation symbols are generated using Gumbel-Max sampling; during backpropagation, a differentiable softmax approximation is used to replace the argmax operation for gradient estimation. General Gumbel-Max sampling methods use independent Gumbel variables to reparameterize discrete variables. Consider a single-hot variable... ,in 1 represents a different category. The classification probability distribution is determined by... It means, set Indicates a Gumbel distribution ,variable It can be by sampling:

[0043] in, index Convert to Form, of which the first One element equals 1, and all other elements equal 0. In order modulation, from right The sampling is represented as:

[0044] in, It is a vector composed of all constellation symbols. It is a random variable with a Gumbel distribution. The aforementioned probability of constellation symbols, one The length of the vector corresponds to the total number of constellation symbols. When the first When the element equals 1, this means the first... The constellation symbols have been sampled.

[0045] To estimate the backpropagation... The gradient of the non-differentiable argmax function is approximated by the differentiable Softmax function. Therefore, the gradient obtained by the Gumbel-Max sampling method... Vectors are Vector substitution, where the first Each element is represented as:

[0046] in, This represents the temperature hyperparameter used to control approximate hardness. Unlike forward propagation, in backward propagation, there is... , No. The formula for calculating each element is:

[0047] Through the above transformation, the loss function becomes differentiable everywhere, and with respect to the parameters. The gradient can be obtained through backpropagation. The samples were obtained. Therefore, The gradient estimate can be expressed as:

[0048] in, Indicates independence The function.

[0049] Step 2: Output Reshaped into , This represents the number of subcarriers in an OFDM symbol. This refers to the number of OFDM data symbols. The reshaped constellation symbol sequence is passed through a power constraint, and the signal is subject to the following average power constraint:

[0050] Among them, expectations It is calculated based on the input image. Represents the Frobenius norm. Without loss of generality, assume... .

[0051] Step 3: Next, use IFFT to convert OFDM symbols into OFDM signals. To address issues such as multipath channels and inter-symbol interference, a CP is inserted into the OFDM signal, resulting in a transmitted frequency-domain complex-valued signal.

[0052] in, It is the CP length. It represents the number of pilot symbols.

[0053] Step 4: Local Pilot Symbols For information known at both the transmitting and receiving ends, used for channel estimation at the receiving end, without loss of generality, this invention assumes that each pilot symbol is identical. The signal received by the receiving end after passing through the wireless channel is represented as:

[0054] in, This represents the convolution operation. This represents additive white Gaussian noise, where each path undergoes independent Rayleigh fading, satisfying... , , This represents the number of multipath paths. The power of each path follows an exponential decay curve. ,in, The normalized coefficients satisfy , This is the time decay constant. Note that, given random channel parameters, the channel transfer function is... It is fully differentiable. This means that when an end-to-end system is trained using the SGD method, the gradient from the semantic decoder can pass through multipath fading channels. Backpropagation is achieved to the semantic code modulator.

[0055] Step 5: The receiver receives the noisy signal output from the channel. First, remove the CP (Pilot Symbol), then use DFT (Digital Transformation Theory) to transform the data to the frequency domain to obtain the pilot symbols. and data symbols .Will and The following channel estimate is obtained by combining the results:

[0056] in, Indicates local pilot signal. It is the first k Channel estimation gain for each subcarrier .

[0057] Step Six: Estimation This will be fed back to the dual attention module in the semantic encoder and semantic decoder neural network structure, where the dual attention module consists of, for example, ... Figure 3 The neural network shown is composed of an average pooling layer, a fully connected layer, a PReLU layer, a fully connected layer, and a sigmoid layer, designed to utilize CSI information for more adaptive carrier allocation. Data symbols after DFT. Using channel estimation results To achieve equilibrium:

[0058] Step 7: The receiving end will receive the equalized data. The data is fed into the semantic decoder to reconstruct the source data and complete the corresponding intelligent task. Specifically, the received data... The semantic decoder estimates the posterior distribution. ,in This represents the neural network parameters. Then, the source data is recovered through maximum a posteriori probability decoding: .

[0059] Step 8: Calculate the loss function and update the gradient; To maximize mutual information, the information-maximizing (Info-Max) method based on information theory is adopted as the design principle of the system framework of this invention. Specifically, the goal of this invention is to be able to receive sequences The goal is to retain as much information as possible from the source data. Therefore, an objective function based on maximizing mutual information is defined, and the optimization problem is established as follows:

[0060] Because high-dimensional variables have complex distributions, it is difficult to estimate their joint and marginal distributions, which makes direct optimization of the objective function of mutual information challenging. Furthermore, this optimization problem only considers the semantic encoder-modulator parameters. The semantic decoder at the receiver needs to be jointly trained with the semantic codec modulator at the transmitter. Therefore, a loss function needs to be designed to take into account the optimization of the semantic decoder. This embodiment designs a general loss function to jointly optimize the transition probability at the transmitter. and receiver probability First, construct the true posterior distribution. An approximate inference model is derived, and then variational inference is used to derive the following lower bound based on the mutual information objective function:

[0061] This lower bound includes the true posterior distribution. With neural network parameterized posterior distribution Distance metric between. Indicates that under given parameters Next, input image With the received equalized signal The statistical correlation between them measures the difference between the receiver and the receiver. How much information can be obtained from this? Information. Indicates the received equalization signal Given a transmitted signal Y The expectation under the conditional distribution simulates the randomness of a signal after it passes through a wireless channel. Indicates the encoder-modulator output F Given an input image and parameters The expectation under the conditional distribution represents the process by which the transmitter generates and sends signals based on semantic information. Indicates the input image Given the received equalization signal The expectation under the posterior distribution reflects the uncertainty of the receiver in inferring semantic information based on the received signal. Indicates the input image The entropy represents The uncertainty or information content is a constant that is independent of system parameters.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A digital semantic communication method that combines learnable constellation mapping and channel adaptation, characterized in that, The method includes the following steps: S1: Input the source data into the semantic coding modulator to obtain a differentiable constellation symbol; S2: Apply average power constraints to the signal using a power constraint unit; S3: The signal is fed into the OFDM transmitter to perform an inverse fast Fourier transform and add a cyclic prefix, mapping it to a baseband signal; S4: Passing through the wireless channel; S5: The signal is fed into the OFDM receiver for discrete Fourier transform and cyclic prefix removal, and channel estimation is performed. S6: Perform equalization using the channel estimation results, and feed the channel estimation results back to the semantic coding modulator and semantic decoder. S7: Reconstruct source data using a semantic decoder; S8: Calculate the loss function to jointly optimize the transition probabilities of the sender and receiver, and update the gradient.

2. The digital semantic communication method according to claim 1, characterized in that, In step S1, the semantic coding modulator includes a probabilistic encoder and a learnable constellation symbol mapping module for processing source data. Generate constellation symbol sequence ; The constellation symbol mapping module uses parameters The neural network is parameterized to learn the transition probabilities; in, This represents a semantic encoding modulator consisting of a probabilistic encoder and a learnable constellation symbol mapping module. It is the source data, i.e., the input image. c , h and w These represent color, height, and width, respectively.

3. The digital semantic communication method according to claim 2, characterized in that, Step S1 specifically includes: for The transition probability model of the first-order modulation, taking a length of... Constellation symbol sequence , , each component Modeled as conditionally independent variables, where the components Choose your zodiac sign of One of the values ​​in a set of symbols, the neural network outputs each symbol. The classification distribution will It's a constellation symbol. The probability of one of them is expressed as: Since it is assumed that the components are conditionally independent, the joint probability mass function of the semantic coding modulator is expressed as: in, Indicates an indicator function, Indicates when yes The probability of one of the symbols in the diagram; Constellation symbols are generated using a reparameterization technique based on Gumbel-Softmax. End-to-end training is achieved through the following mechanism: during forward propagation, discrete constellation symbols are generated using Gumbel-Max sampling, and during backpropagation, gradient estimation is performed by replacing the argmax operation with a differentiable softmax approximation.

4. The digital semantic communication method according to claim 3, characterized in that, Step S2 specifically includes: outputting the constellation symbol sequence. Reshaped into , This represents the number of subcarriers in an OFDM data symbol. This refers to the number of OFDM data symbols; the reshaped constellation symbol sequence is passed through a power constraint, and the signal is subject to the following average power constraint: Among them, expectations It is calculated based on the input image. Represents the Frobenius norm; This indicates an average power constraint.

5. The digital semantic communication method according to claim 4, characterized in that, Step S3 specifically includes: converting OFDM symbols into OFDM signals using inverse fast Fourier transform; inserting CP into the OFDM signal to form the transmitted frequency domain complex-valued signal. , represented as: in, It is the CP length. It represents the number of pilot symbols.

6. The digital semantic communication method according to claim 5, characterized in that, In step S4, after passing through the wireless channel, the signal received by the receiving end... Represented as: in, Indicates the wireless channel. Indicates the convolution operation; This represents additive white Gaussian noise, where each path undergoes independent Rayleigh fading, satisfying... , , For the number of multipaths, Indicates the first l A path.

7. The digital semantic communication method according to claim 6, characterized in that, In step S5, the receiving end receives the noisy signal output from the channel. First, the cyclic prefix is ​​removed, and then the data is transformed to the frequency domain using Discrete Fourier Transform to obtain the pilot symbols. and data symbols ;Will and The following channel estimation is obtained jointly: in, This represents the channel estimation result. Indicates the local pilot signal; It is the first k Channel estimation gain for each subcarrier .

8. The digital semantic communication method according to claim 7, characterized in that, In step S6, the data symbols after discrete Fourier transform Using channel estimation results Achieve the following equilibrium: The receiving end will equalize the data The data is fed into the semantic decoder to reconstruct the source data and complete the corresponding intelligent task; simultaneously, the estimated... The feedback will be fed into the dual attention module in the semantic encoder-modulator and semantic decoder, where the dual attention module consists of two neural network architectures; each consists of an average pooling layer, a fully connected layer, a PReLU layer, a fully connected layer, and a Sigmoid layer.

9. The digital semantic communication method according to claim 8, characterized in that, Step S7 specifically includes: receiving The semantic decoder estimates the posterior distribution. Then, the source data is recovered by decoding with maximum a posteriori probability: ,in This represents the parameters of the neural network.

10. The digital semantic communication method according to claim 9, characterized in that, In step S8, an objective function based on maximizing mutual information is defined, and the optimization problem is established as follows: Design a general loss function to jointly optimize the transition probability at the sender. and receiver probability First, construct the true posterior distribution. An approximate inference model is derived, and then variational inference is used to derive the following lower bound based on the mutual information loss function: This lower bound includes the true posterior distribution. With neural network parameterized posterior distribution The distance metric between them; where, Indicates that under given parameters Next, input image With the received equalized signal Statistical correlation between them; Indicates the received equalization signal Given a transmitted signal Y The expectation under the conditional distribution; Indicates the encoder-modulator output F Given an input image and parameters The expectation under the conditional distribution; Indicates the input image Given the received equalization signal The expectation under the posterior distribution; Indicates the input image The entropy.

Images