A semantic communication method and system based on interleaved frequency division multiplexing

Abstract

The application discloses a semantic communication method and system based on interleaved frequency division multiplexing, belongs to the field of semantic communication, and is applied to a semantic communication system. The method comprises the following steps: a sending end extracts semantic information through a semantic encoder and transmits the semantic information after modulation through interleaved frequency division multiplexing; a receiving end reconstructs a semantic signal based on a received signal through a cross-domain iterative detector and completes a target task through a semantic decoder. The cross-domain iterative detector performs linear detection in a time domain by utilizing channel sparsity, performs nonlinear detection in a characteristic domain based on a fractional model, and realizes cross-domain transformation through an interleaved frequency division multiplexing modulation matrix with universal characteristics. The application can realize training of an encoder / decoder and decoupling of a channel through interleaved frequency division multiplexing modulation, the channel has strong universality, and the compatibility with an existing digital communication system is good. Since the type of the encoder / decoder is not limited, the whole system can perform multiple tasks, and the task universality is good.

Description

Technical Field

[0001] This invention belongs to the field of semantic communication technology, specifically relating to a semantic communication method and system based on interleaved frequency division multiplexing. Background Technology

[0002] With the explosive growth of 6G mobile communication and intelligent applications such as Extended Reality (XR), autonomous driving, and the metaverse, wireless networks face the challenge of transmitting massive amounts of heterogeneous data (such as text, images, and video), placing extremely high demands on low latency and high reliability. Traditional 5G communication systems focus on ensuring accurate bit-level transmission, neglecting the semantic content and importance of the information itself. This leads to a huge waste of bandwidth resources and makes it difficult to meet the efficient transmission needs of future intelligent services. To solve this problem, Semantic Communication (SemCom) has emerged as a new paradigm. Unlike traditional communication that focuses on bit transmission, semantic communication uses deep learning technology to extract and transmit task-related compact semantic features, significantly improving communication efficiency.

[0003] However, current semantic communication research largely focuses on end-to-end (E2E) deep neural network (DNN) architectures. While this architecture performs well in specific channels, its robustness in complex dynamic channels is limited. Furthermore, it replaces traditional source-channel coding modules, leading to serious compatibility issues with current standardized digital communication physical layer infrastructure. This significantly restricts the practical implementation and deployment of semantic communication. Traditional digital-semantic hybrid architectures mostly employ a separate design, making it difficult to guarantee optimal transmission at the physical layer. Summary of the Invention

[0004] This invention provides a semantic communication method and system based on interleaved frequency division multiplexing, which can solve the problems of poor channel universality, semantic task universality, and compatibility with existing digital communication systems in traditional semantic communication methods.

[0005] In a first aspect, embodiments of the present invention provide a semantic communication method based on interleaved frequency division multiplexing, the method being applied to a semantic communication system including a transmitter and a receiver, the method comprising:

[0006] The transmitting end encodes the source data based on the semantic encoder to extract the semantic information required by the target task to obtain the semantic signal, and performs interleaved frequency division multiplexing modulation on the semantic signal to obtain the time domain signal, wherein the interleaved frequency division multiplexing modulation matrix has universal characteristics.

[0007] The transmitting end transmits the time-domain signal;

[0008] The receiving end is based on a cross-domain iterative detector and uses the received signal to perform iterative updates in order to recover semantic information and obtain a semantic reconstruction signal;

[0009] The cross-domain iterative detector is used to: perform linear detection based on the received signal in the time domain using channel sparsity, perform nonlinear detection based on a fractional model in the feature domain, and realize cross-domain transformation of the signal between the time domain and the feature domain based on the interleaved frequency division multiplexing modulation matrix.

[0010] The receiving end inputs the semantic reconstruction signal to the semantic decoder to complete the target task.

[0011] Secondly, embodiments of the present invention provide a semantic communication system based on interleaved frequency division multiplexing, including a transmitter and a receiver;

[0012] The sending end is used for:

[0013] The source data is encoded based on a semantic encoder to extract the semantic information required by the target task and obtain a semantic signal. The semantic signal is then interleaved with frequency division multiplexing modulation to obtain a time-domain signal. The interleaved frequency division multiplexing modulation matrix has universal characteristics.

[0014] Transmit the time-domain signal;

[0015] The receiving end is used for:

[0016] Based on a cross-domain iterative detector, the received signal is used for iterative updates to recover semantic information and obtain a semantic reconstruction signal;

[0017] The cross-domain iterative detector is used to: perform linear detection based on the received signal in the time domain using channel sparsity, perform nonlinear detection based on a fractional model in the feature domain, and realize cross-domain transformation of the signal between the time domain and the feature domain based on the interleaved frequency division multiplexing modulation matrix.

[0018] The semantic reconstruction signal is input into the trained semantic decoder to complete the target task.

[0019] The beneficial effects of this invention compared to existing technologies are as follows: Since this invention transmits signals after interleaving frequency division multiplexing modulation, and the modulation matrix used during modulation has universal characteristics, it achieves statistical decoupling of semantic features and the physical channel. This makes the training of the semantic encoder / decoder independent of channel conditions, avoiding the need for retraining when channel conditions change. This allows the system to adapt to complex and dynamic channel environments, achieving decoupling of encoder / decoder training from channel conditions, improving channel universality and compatibility with existing digital communication systems. Furthermore, because it does not limit the type of encoder / decoder, the entire system can perform multiple tasks, and this invention also improves the versatility of the semantic tasks performed. In addition, by updating semantic information between the time domain and the feature domain, the sparsity of the time-domain signal can be utilized to reduce computational complexity. Moreover, due to the universal characteristics of the modulation matrix, the variance of the signal remains unchanged when transforming between the two domains, further reducing computational complexity and thus lowering costs. Attached Figure Description

[0020] Figure 1 This invention provides a schematic diagram of the structure of a semantic communication system based on interleaved frequency division multiplexing in a single-input single-output scenario.

[0021] Figure 2 This invention provides a schematic diagram of the structure of a semantic communication system based on interleaved frequency division multiplexing in a multi-input multi-output scenario.

[0022] Figure 3 This is a schematic diagram of the structure of a cross-domain iterative detector provided in an embodiment of the present invention;

[0023] Figure 4 A schematic diagram illustrating the implementation process of a semantic information iterative update method provided in an embodiment of the present invention;

[0024] Figure 5 A flowchart illustrating the implementation of a training method for a semantic encoder, decoder, and score model provided in an embodiment of the present invention;

[0025] Figure 6 A flowchart illustrating the implementation of a semantic communication method based on interleaved frequency division multiplexing provided in an embodiment of the present invention;

[0026] Figure 7 A comparison chart of PSNR indices for different modulation methods in a SISO scenario is provided as an embodiment of the present invention.

[0027] Figure 8 A comparison chart of PSNR indices for different modulation methods in a MIMO scenario is provided as an embodiment of the present invention.

[0028] Figure 9This invention provides a comparison chart of SSIM metrics for different modulation methods in a MIMO scenario, as shown in an embodiment of the invention.

[0029] Figure 10 This is a schematic diagram showing the PSNR index variation curve of different semantic communication methods as a function of SNR, and the performance prediction curve of the present invention, provided for embodiments of the present invention. Detailed Implementation

[0030] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0031] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0032] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0033] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0034] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0035] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0036] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0037] Figure 1 The diagram shown is a schematic of the structure of a semantic communication system based on interleaved frequency division multiplexing in a single-input single-output (SISO) scenario provided by the present invention.

[0038] For example, the system may consist of a transmitter and a receiver.

[0039] As an example, see Figure 1 In the SISO scenario, at the transmitting end, the input source data, such as images, text, and videos, can first be processed by a semantic encoder to extract task-related high-dimensional semantic features, which are then mapped into complex semantic vectors (also known as semantic signals) that satisfy power constraints. Subsequently, the semantic vectors are interleaved with frequency division multiplexing modulation to generate a time-domain signal. Afterward, to combat multipath effects, a cyclic prefix (CP) is added to the time-domain signal, and it is transmitted through a transmit filter.

[0040] For example, semantic signals can satisfy the following formula:

[0041] ,

[0042] in, For semantic signals, express It is a length of A one-dimensional complex vector, For source data, This indicates the feature extraction and mapping operations performed by the semantic encoder.

[0043] Specifically, the power constraint satisfied by the semantic vector is: ,in, This indicates taking the modulus.

[0044] In one possible implementation, the interleaved frequency division multiplexing modulation matrix has universal properties.

[0045] Here, "universality class" is not an adjective like "relevant" or "universal," but rather a general, proprietary term. Specifically, if the equivalent channel matrix... If it belongs to the universal class of matrices, then For signal Its universal reusability.

[0046] in, Given a channel matrix, This represents the interleaved frequency division multiplexing modulation matrix. It is a uniform random phase matrix. Indicates diagonalization, ,express to obey Uniform distribution express to . It is a definite spectral convergent matrix with a bounded spectral norm. , Furthermore, for any fixed , , satisfy: , Represents the set of positive integers. Represents the trace of a matrix. Denotes the maximum norm. This indicates asymptotic less than or equal to.

[0047] Optionally, It can also be an affine diagonal matrix, satisfying: ,in, This is the affine parameter. The selection of this parameter can be based on the formula: Determined based on the maximum Doppler spread and delay spread of the channel. For the normalized maximum Doppler extension, It is the Doppler interval factor; it can also be selected in other forms according to the time delay spread.

[0048] In one example, the interleaved frequency division multiplexing modulation matrix can be a random unitary matrix with universal class properties.

[0049] For example, a time-domain signal can be represented as:

[0050] ,

[0051] in, For time-domain signals, Let be an interleaved frequency division multiplexing modulation matrix, and satisfy... , It is a random interleaved matrix. ,or , It is a random permutation matrix; For signal length, This indicates the conjugate transpose. This represents a unitary transform matrix, such as the Inverse Fast Fourier Transform (IFFT), Discrete Cosine Transform (DCT), Walsh-Hadamard Transform (WHT) matrix, etc.; or an interleaved frequency division multiplexing modulation matrix. It can also be a Haar matrix.

[0052] Optionally, the coverage of the interleaved frequency division multiplexing modulation matrix may include both the real and complex domains.

[0053] As an example, see Figure 1 At the receiving end, the time-domain signal is first received and the cyclic prefix is ​​removed. Then, the received signal is input to a cross-domain iterative detector, which uses the received signal to iteratively update the semantic information, obtaining a semantically reconstructed signal. Finally, the semantically reconstructed signal is input to a trained semantic decoder to complete the target task.

[0054] For example, the target task can be to recover the source data, or it can be a task such as classification, detection, or recognition.

[0055] In one possible implementation, during the iterative update process, a cross-domain iterative detector is used to perform linear detection processing based on the received signal in the time domain by utilizing the channel sparsity, and to perform nonlinear detection processing based on a fractional model in the feature domain; at the same time, the cross-domain transformation of the signal between the time domain and the feature domain is realized based on the interleaved frequency division multiplexing modulation matrix.

[0056] Specifically, since the interleaved frequency division multiplexing modulation matrix has universal characteristics, regardless of whether the physical channel is additive white Gaussian noise, Rayleigh fading or other types, it tends to be consistent in a statistical sense after being transformed by the interleaved frequency division multiplexing modulation matrix. The end-to-end transmission characteristics experienced by the semantic information depend only on the random distribution of the interleaved frequency division multiplexing modulation matrix, and are not related to the specific implementation of the underlying physical channel.

[0057] In one example, the received signal after filtering and removing the CP can be represented as:

[0058] ,

[0059] in, Indicates receiving signal, The time-domain channel matrix, It is additive white Gaussian noise.

[0060] Because this invention transmits signals after interleaving frequency division multiplexing modulation, and the modulation matrix used during modulation has universal characteristics, it achieves statistical decoupling between semantic features and the physical channel. This makes the training of the semantic encoder / decoder independent of channel conditions, avoiding the need for retraining when channel conditions change. This allows the system to adapt to complex and dynamic channel environments, decoupling encoder / decoder training from channel conditions and improving channel universality and compatibility with existing digital communication systems. Since it does not restrict the type of encoder / decoder, the entire system can perform multiple tasks, further improving the versatility of the semantic tasks performed. Furthermore, by updating semantic information between the time domain and the feature domain, the sparsity of the time-domain signal can be utilized to reduce computational complexity. And due to the universal characteristics of the modulation matrix, the variance of the signal remains unchanged when transforming between the two domains, further reducing computational complexity and thus lowering costs.

[0061] Figure 2 The diagram shown is a schematic of the structure of a semantic communication system based on interleaved frequency division multiplexing in a multiple-input multiple-output (MIMO) scenario provided by the present invention.

[0062] As an example, to further improve the system's spectral efficiency and transmission reliability, the system can also be extended to MIMO application scenarios.

[0063] Similarly, in the MIMO scenario, the system also includes a receiver and a transmitter, and the two devices perform the same functions as in the SISO scenario. The difference lies in the number of antennas, the time domain signal, and the composition of the received signal at the transmitter and receiver.

[0064] For example, if the sending end is configured with One transmitting antenna, the receiving end is equipped with The first receiving antenna, the first transmitting antenna. The time-domain signal transmitted by the transmitting antenna can be expressed as: , Less than or equal to If the integer is a positive integer, then the receiving end's first... Received signal on root receiving antenna , Less than or equal to positive integers, express Also of length A one-dimensional complex vector; can be represented as the superposition of the time-domain signals from all transmitting antennas after fading along their respective paths, satisfying the following formula:

[0065] ,

[0066] in, Indicates the first root transmitting antenna to the first The time-domain channel matrix between the root receiving antennas, For the first The additive white Gaussian noise vector at the root receiving antenna.

[0067] At this point, the input-output relationship of the entire system in a MIMO scenario can be represented as a block matrix:

[0068] ,

[0069] in, In order to receive signals, Indicates matrix transpose. This is the time-domain channel matrix, containing all sub-channel matrices. , It is a time-domain signal.

[0070] Figure 3 The diagram shown illustrates the structure of a cross-domain iterative detector according to an embodiment of the present invention. As an example and not a limitation, the cross-domain iterative detector may include a time-domain linear detector, a feature-domain nonlinear detector, and a cross-domain transformation module.

[0071] In some embodiments, during each iteration update of the cross-domain iterative detector, the received signal can first be input to the time-domain linear detector, which utilizes the sparsity of the time-domain channel to perform linear detection based on the received signal, obtaining a time-domain orthogonal signal. Then, the cross-domain transformation module performs a cross-domain transformation on the time-domain orthogonal signal based on the interleaved frequency division multiplexing modulation matrix to demodulate and obtain the input signal for nonlinear detection. Next, the nonlinear detector in the feature domain performs nonlinear detection processing based on a fractional model to obtain a feature-domain orthogonal signal. Finally, it is determined whether a preset requirement is met. If not, the cross-domain transformation module performs an inverse cross-domain transformation on the feature-domain orthogonal signal based on the interleaved frequency division multiplexing modulation matrix to achieve modulation, and proceeds to the next iteration. If the requirement is met, the signal after feature-domain nonlinear detection but before orthogonal processing (i.e., the posterior output of the nonlinear detection described below) is output as the semantic reconstruction signal.

[0072] In one possible implementation, see Figure 3 A time-domain linear detector may include a damping module, a memory matched filter, and a first orthogonal module; a feature-domain nonlinear detector may include a linear estimator and a second orthogonal module.

[0073] In one example, see Figure 4 The cross-domain iterative detector can specifically perform iterative updates to recover semantic information through the following steps S401~S409.

[0074] S401, the first The signal before damping in the next iteration is input to the damping module for damping operation, resulting in the signal before damping. The damped signal after the next iteration.

[0075] For example, the first The damped signal after the next iteration can be expressed as: It is the estimated signal. The combination of .

[0076] In one example, when When equal to 1, initialize ;when When it is greater than 1, it constitutes The estimated signal can satisfy the following formula:

[0077] ,

[0078] in, , The first , An estimated signal, For the first The signal before damping in the next iteration. It is a positive integer; Let be the damping vector. , For the first The two damping coefficients used in the next iteration update.

[0079] For example, the damping coefficient can be based on The variance is used to calculate and optimize convergence performance.

[0080] Specifically, the damping coefficient can satisfy the following formula:

[0081] ,

[0082] ,

[0083] in, , They are respectively , No. Signal before damping in the next iteration The variance.

[0084] S402, the first The damped signal and the received signal from the next iteration are input to a memory matched filter for linear detection, yielding the linear detection result. The posterior output of the next iteration.

[0085] In one example, the memory matched filter can be a low-complexity memory matched filter based on the previous iterative estimation results (damped signal), expressed as: ,in, Indicates the first The memory matched filter at the next iteration Represents the memory matched filter for the first The damped signal of the nth iteration is subjected to linear detection processing, i.e., the nth iteration... The posterior output of the next iteration.

[0086] Specifically, the process by which the memory matched filter processes the damped signal based on the received signal can be expressed as follows:

[0087] ,

[0088] in:

[0089] ,

[0090] ,

[0091] in, This is a scaling factor used to accelerate the convergence speed of the time-domain linear detector. For the first The memory matched filter at the next iteration For the first The damped signal after the next iteration It is also a scaling factor, used to accelerate convergence. , They represent The maximum and minimum eigenvalues.

[0092] Similarly, initialization .

[0093] S403, the first The damped signal after the nth iteration, the linearly detected signal The posterior output of the nth iteration is input into the first orthogonal module for orthogonalization processing to obtain the nth iteration. The signal after time-domain orthogonality in the next iteration.

[0094] For example, the first... The time-domain orthogonal signals after the next iteration can satisfy the following formula:

[0095] ,

[0096] in, For the first The time-domain orthogonal signal after the next iteration Indicates the first The time-domain linear detector at the next iteration , These are the normalized coefficients and orthogonality coefficients of the time-domain linear detector, respectively. Values.

[0097] S404, the first The time-domain orthogonalized signal from the next iteration is input to the cross-domain transformation module for demodulation based on the interleaved frequency division multiplexing modulation matrix, yielding the nonlinear detection result. The input signal for the next iteration.

[0098] For example, the first nonlinear detection The input signal and its variance for each iteration can satisfy the following formulas:

[0099] ,

[0100] ,

[0101] in, The first nonlinear detection The input signal for the next iteration This represents the inverse matrix. Depend on variance constitute, for The variance.

[0102] Due to the universality of the interleaved frequency division multiplexing modulation matrix, or in other words, due to the unitary matrix property of the interleaved frequency division multiplexing modulation matrix, the (inverse) cross-domain transformation of the signal between the time domain and the characteristic domain does not change the variance. .

[0103] S405, the first nonlinear detection The input signal of the nth iteration is input to the nonlinear estimator for nonlinear detection, and the result is the nth iteration of the nonlinear detection. The posterior output of the next iteration.

[0104] In one possible implementation, since the prior distribution of semantic signals is unknown and complex, traditional Bayesian estimation methods fail. This invention uses a fraction-based generative model to approximate the posterior estimation.

[0105] In one example, the nonlinear estimator could be a denoising estimator based on Tweedie's Formula. Specifically, it could be based on a pre-trained score model to approximate the... The gradient of the marginal log-likelihood probability (i.e., the fractional function) is used to avoid explicitly solving for the prior distribution of the semantic signal. Then, based on the estimated gradient, variance Denoising estimation is performed to obtain the first nonlinear detection. The posterior output of the next iteration.

[0106] For example, the first nonlinear detection The posterior output of the next iteration can satisfy the following formula:

[0107] ,

[0108] in, Indicates the first Nonlinear estimator at the next iteration The first linear detection The posterior output of the next iteration. Representation of fractional models The estimated gradient.

[0109] In one example The variance can be updated using the Stein's Unbiased Risk Estimate (SURE) algorithm.

[0110] For example, updated via SURE The variance can satisfy the following formula:

[0111] ,

[0112] in, The first nonlinear detection The variance of the input signal in the next iteration For signal length, express variance Represents a nonlinear estimator right Find the partial derivative. To find the sign of the partial derivative, Indicates the element index.

[0113] Specifically, the Jacobian term It can be estimated using sampling methods.

[0114] In another example, an additional fractional model (called a second-order fractional model) can be trained to approximate the trace of the Hessian matrix of the fractional function. This term contains second-order information about the prior distribution of the semantic signal, and thus can be combined with the second-order generalization estimate of Tweedie's Formula in this way. The variance.

[0115] In yet another example, residual calculation can also be used. The variance.

[0116] For example, calculated via residuals The variance can satisfy the following formula:

[0117] ,

[0118] in, The first nonlinear detection The variance of the input signal in the next iteration express The square of the Frobenius norm, Indicates received signal Total noise energy in This represents the total number of dimensions of the noise.

[0119] In another possible implementation, other types of generative models, such as diffusion models, variational autoencoders, and generative adversarial networks, can be used to implicitly or explicitly model the prior distribution of semantic signals.

[0120] S406, Determine whether the preset requirements are met.

[0121] For example, it can be determined by... Whether the value is greater than or equal to the threshold determines whether the preset requirements are met.

[0122] In one example, if the preset requirements are met, step S407 can be performed; if not, steps S408 and S409 can be performed sequentially, and then... The next iteration update will begin from step S401.

[0123] S407, the first nonlinear detection The posterior output of the next iteration is used as the semantic reconstruction signal output.

[0124] S408 will perform the nonlinear detection of the first... The input signal of the next iteration, the first nonlinear detection The posterior output of the nth iteration is input into the second orthogonal module for orthogonalization processing to obtain the nth iteration. The signal after the feature domains of the next iteration are orthogonal.

[0125] Orthogonalization ensures that the estimation error of the output is minimized. / ) and the estimation error of the input ( / The signals are orthogonal, thus eliminating the correlation between update messages. This also ensures that the nonlinear detection input error is independent and identically distributed Gaussian, and is independent of the true time-domain signal.

[0126] For example, the first The signal after the feature domains of the next iteration are orthogonalized satisfies the following formula:

[0127] ,

[0128] in, For the first The signal after the feature domains of the next iteration are orthogonal. This indicates a nonlinear detector. The first normalized coefficient used for the nonlinear detector One value, The first orthogonal coefficient used for the second normalization module Values.

[0129] S409, the first The signal after orthogonalization of the feature domains in the next iteration is input to the cross-domain transformation module to perform inverse cross-domain transformation based on the interleaved frequency division multiplexing modulation matrix, thereby achieving modulation and obtaining the first... The signal before damping in the next iteration.

[0130] For example, the first The signal and its variance before damping in the next iteration can satisfy the following formula:

[0131] ,

[0132] ,

[0133] in, For the first The signal before damping in the next iteration. Depend on variance constitute, for Dimensional unit matrix.

[0134] If traditional message passing detection algorithms are directly applied to the received signal obtained after interleaved frequency division multiplexing modulation, there will be a problem of high computational complexity. The cross-domain iterative detector provided by this invention combines a fraction-based generative model to process the unknown prior distribution of high-dimensional semantic features. It can make full use of the sparsity of the time-domain channel and the universality of the interleaved frequency division multiplexing modulation matrix to achieve highly reliable semantic signal recovery with low computational complexity.

[0135] This invention also provides a method for asymptotic performance analysis of cross-domain iterative detectors:

[0136] Under the condition of full orthogonality, if the estimation error of the cross-domain iterative detector is to be asymptotically independent and identically distributed (IID) Gaussian, then the estimation error of the cross-domain iterative detector must satisfy:

[0137] ,

[0138] ,

[0139] in, , , , They are both columnar IID Gaussian and row-based joint Gaussian, and both are related to the time-domain signal. Irrelevant , For respectively the first The estimation errors in the feature domain and time domain during the next iteration. Less than or equal to Positive integers.

[0140] Because the interleaved frequency division multiplexing modulation matrix has universal class properties, or in other words, unitary matrix properties, the cross-domain transform and inverse cross-domain transform do not change the variance. Therefore, the mean square error function can be used to accurately predict the asymptotic performance of the cross-domain iterative detector, i.e.:

[0141] ,

[0142] ,

[0143] in, , Let represent the covariance matrices of the estimation errors in the time domain and the feature domain, respectively. , They represent , The first in the matrix Values. , Let be the mean square error functions for linear detection and nonlinear detection, respectively. Less than or equal to Positive integers.

[0144] Figure 5 The diagram illustrates a training method for a semantic encoder, decoder, and score model provided by an embodiment of the present invention. As an example and not a limitation, this method can be applied to a receiving / transmitting end or any other electronic device. After training, the semantic encoder is transferred to the transmitting end, and the semantic decoder and score model are transferred to the receiving end. The method may include steps S501 to S503, which are described below.

[0145] S501 is used for joint training of the semantic encoder and semantic decoder.

[0146] In one example, since the semantic communication method provided by this invention does not depend on a specific channel model, the optimization objective of the semantic encoder / decoder can be set to minimize task-specific loss, such as minimizing the mean square error of the reconstruction task or minimizing the cross-entropy loss of the classification task.

[0147] For example, the optimization objective of a semantic encoder / decoder can be expressed as:

[0148] ,

[0149] in, Indicates minimization. Encoder network parameters For decoder network parameters, Expressing expectations, Indicates source data, This indicates that the source data was selected from the dataset. This represents the feature extraction and mapping operations of the semantic encoder. This represents the decoding operation of the semantic decoder. This represents the loss between the source data and the source data recovered after semantic encoding and decoding.

[0150] In this step, training focuses solely on improving task performance.

[0151] S502, continuously injects noise into the sample semantic signal until the sample semantic signal is submerged by noise, thus obtaining the noisy semantic signal.

[0152] In one possible implementation, the fractional model can be trained within the framework of variance explosion stochastic differential equations. The training process of the fractional model includes a forward process and a backward process. The forward process refers to step S502, which involves injecting noise into the clean semantic signal generated by the trained semantic encoder.

[0153] In one example, the noise injection process can be described by stochastic differential equations:

[0154] ,

[0155] in, For diffusion time, , Indicates diffusion time as The semantic signal at that time, which contains noise. The variance is , This is the endpoint of the diffusion process, at which point the semantic signal is completely submerged by noise.

[0156] S503 trains the fractional model based on noisy semantic signals.

[0157] In one possible implementation, the inverse process can utilize a fractional model to denoise the noisy semantic signal based on the constants of the inverse differential equation, so that it learns the prior distribution information of the semantic signal required for the signal denoising process.

[0158] In one example, the inverse denoising learning process can be described using stochastic differential equations:

[0159] ,

[0160] in, This is the fractional function term that the fractional model needs to learn, which contains prior distribution information of the semantic signals required by the cross-domain iterative detector.

[0161] For example, the loss function used in the reverse process The following formula can be satisfied:

[0162] ,

[0163] in, Represents the L2 norm. It is a weighting coefficient.

[0164] The trained score model Prior distribution information of semantic signals can be provided in cross-domain iterative detectors to assist in signal distribution.

[0165] Figure 6 The diagram shown illustrates a semantic communication method based on interleaved frequency division multiplexing provided by an embodiment of the present invention. This method can be applied to the aforementioned system and may include steps S601-S605, which are described below.

[0166] S601, the sending end inputs the source data into the trained semantic encoder to extract the semantic information required by the target task and obtain the semantic signal.

[0167] For example, a semantic encoder can extract high-dimensional semantic features from source data and map them into semantic signals that satisfy power constraints.

[0168] S602, the transmitting end performs interleaved frequency division multiplexing modulation on the semantic signal to obtain the time domain signal.

[0169] For example, the interleaved frequency division multiplexing modulation matrix has universal properties.

[0170] S603, the transmitting end transmits time-domain signals.

[0171] Accordingly, the receiving end receives the time-domain signal, filters it, and removes the CP (Concurrent Phase) to obtain the received signal.

[0172] S604, the receiver is based on a cross-domain iterative detector, which uses the received signal to perform iterative updates in order to recover semantic information and obtain a semantic reconstruction signal.

[0173] Optionally, the receiving end can also recover semantic information based on iterative detection algorithms or linear detection algorithms such as expectation propagation detection algorithms, approximate message passing detection algorithms, generalized approximate message passing algorithms, and variational message passing algorithms.

[0174] S605, the receiver inputs the semantic reconstruction signal into the trained semantic decoder to complete the target task.

[0175] Because this invention transmits signals after interleaving frequency division multiplexing modulation, and the modulation matrix used during modulation has universal characteristics, it achieves statistical decoupling between semantic features and the physical channel. This makes the training of the semantic encoder / decoder independent of channel conditions, avoiding the need for retraining when channel conditions change. This allows the system to adapt to complex and dynamic channel environments, decoupling encoder / decoder training from channel conditions and improving channel universality and compatibility with existing digital communication systems. Since it does not restrict the type of encoder / decoder, the entire system can perform multiple tasks, further improving the versatility of the semantic tasks performed. Furthermore, by updating semantic information between the time domain and the feature domain, the sparsity of the time-domain signal can be utilized to reduce computational complexity. And due to the universal characteristics of the modulation matrix, the variance of the signal remains unchanged when transforming between the two domains, further reducing computational complexity and thus lowering costs.

[0176] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0177] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0178] To better illustrate the beneficial effects of the present invention, the following simulation experiment was conducted.

[0179] Simulation Experiment 1

[0180] For example, Experiment 1 compares the method provided in this invention (denoted as AE+IFDM+CD-S-MAMP) with a method that replaces the IFDM modulation in this invention with traditional Orthogonal Frequency Division Multiplexing (OFDM) modulation (denoted as AE+OFDM+CD-S-MAMP), a method that replaces the cross-domain iterative detector in this invention with traditional Linear Minimum Mean Square Error (LMMSE) detection (denoted as AE+IFDM+LMMSE), and a detection method that replaces the IFDM modulation in this invention with OFDM and replaces the cross-domain iterative detector with LMMSE (denoted as AE+OFDM+LMMSE). The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of each method are evaluated across the entire signal-to-noise ratio (SNR) range.

[0181] See Figure 7 The curves showing the PSNR index as a function of SNR for various methods in the SISO scenario are shown. Figure 8 The curves showing the PSNR as a function of SNR for different modulation methods in the MIMO scenario are shown. Figure 9The curves showing the SSIM index versus SNR for different modulation methods in MIMO scenarios are illustrated. It can be seen that, across the entire SNR range, the peak SNR and structural similarity index of this invention are superior to the OFDM scheme. Particularly under high SNR conditions (SNR = 20 dB), the PSNR of the AE+IFDM+CD-S-MAMP method reaches 27.21 dB, a performance gain of approximately 3.28 dB compared to 23.93 dB for AE+OFDM+CD-S-MAMP. Simultaneously, the SSIM index also improves from 0.76 to 0.8002. In MIMO scenarios, the robustness and anti-interference advantages of this invention are even more pronounced. At SNR = 20 dB, the PSNR of this invention reaches 26.64 dB, achieving a significant gain of 6.61 dB compared to the MIMO-OFDM scheme (20.03 dB) under the same conditions, while the structural similarity index is significantly improved by 0.1035. Simulation results confirm that the present invention has higher signal reconstruction quality and image transmission fidelity in both SISO and MIMO scenarios.

[0182] Simulation Experiment 2

[0183] Experiment 2 simulated the convergence time of the method provided by this invention. The experiment showed that the method provided by this invention can converge within 10 iterations, achieving performance consistent with state evolution prediction. This confirms that it achieves the optimal performance of empirical Bayes with low complexity.

[0184] This is because the time-domain linear detector in the cross-domain iterative detector provided by this invention can make full use of the sparsity of the time-domain channel to reduce computational complexity, while the computational complexity of the feature-domain nonlinear detector mainly depends on the fractional model. Since the dimension of the semantic signal is much lower than the dimension of the image data, which is the main application scenario of the fractional model, the computational complexity of the fractional model here is less than its computational complexity in the image scenario.

[0185] Simulation Experiment 3

[0186] Experiment 3 compares the method provided by this invention with the traditional deep joint source-channel coding method (denoted as DeepJSCC), and uses the state evolution method to predict the performance of the method provided by this invention.

[0187] Figure 10 The figures shown are the PSNR index variation curves of different methods with SNR and the performance prediction curve of the method provided by the present invention. Figure 10As can be seen from the SE curve, the peak signal-to-noise ratio (PSNR) of the method provided by this invention is superior to that of traditional methods, especially under low SNR conditions, where the gain of this invention exceeds that of traditional methods by approximately 50%. Furthermore, the SE curve also demonstrates that the performance of the method provided by this invention can be accurately predicted using the state evolution method.

[0188] Therefore, since this invention transmits signals after interleaving frequency division multiplexing modulation, and the modulation matrix used during modulation has universal characteristics, it can achieve statistical decoupling between semantic features and the physical channel. This makes the training of the semantic encoder / decoder independent of channel conditions, avoiding the need for retraining when channel conditions change. This allows the system to adapt to complex and dynamic channel environments, achieving decoupling between encoder / decoder training and channel conditions, improving channel universality and compatibility with existing digital communication systems. Because it does not restrict the type of encoder / decoder, the entire system can perform multiple tasks, further improving the versatility of the semantic tasks performed. Furthermore, by updating semantic information between the time domain and the feature domain, the sparsity of the time-domain signal can be utilized to reduce computational complexity. And due to the universal characteristics of the modulation matrix, the variance of the signal remains unchanged when transforming between the two domains, further reducing computational complexity and thus lowering costs.

Claims

1. A method of semantic communication based on interleaved frequency division multiplexing, characterized in that, The method is applied to a semantic communication system based on interleaved frequency division multiplexing, the semantic communication system including a transmitter and a receiver, the method comprising: The transmitting end encodes the source data based on the semantic encoder to extract the semantic information required by the target task to obtain the semantic signal, and performs interleaved frequency division multiplexing modulation on the semantic signal to obtain the time domain signal, wherein the interleaved frequency division multiplexing modulation matrix has universal characteristics. The transmitting end transmits the time-domain signal; The receiving end is based on a cross-domain iterative detector and uses the received signal to perform iterative updates in order to recover semantic information and obtain a semantic reconstruction signal; The cross-domain iterative detector includes a time-domain linear detector, a feature-domain nonlinear detector, and a cross-domain transformation module. The time-domain linear detector includes a damping module, a memory matched filter, and a first orthogonal module. The nonlinear detector includes a nonlinear estimator and a second orthogonal module. The cross-domain iterative detector is used to: perform linear detection based on the received signal in the time domain using channel sparsity, perform nonlinear detection based on a fractional model in the feature domain, and realize cross-domain transformation of the signal between the time domain and the feature domain based on the interleaved frequency division multiplexing modulation matrix. The receiving end uses the received signal to perform the first... The process of each iteration update includes: The first Signal before damping in the next iteration The input is sent to the damping module for damping operation to obtain the first... The damped signal after the next iteration ,in, It is a positive integer; The first The damped signal after the next iteration and the received signal The input is fed into the memory matched filter for linear detection, and the first linear detection result is obtained. The posterior output of the next iteration ; The first The damped signal after the next iteration The first linear detection The posterior output of the next iteration The input is fed into the first orthogonal module for orthogonalization processing to obtain the first... The signal after time-domain orthogonality in the next iteration ; The first The signal after time-domain orthogonality in the next iteration The input is fed into the cross-domain transformation module to perform cross-domain transformation based on the interleaved frequency division multiplexing modulation matrix, thereby achieving demodulation and obtaining the first nonlinear detection result. Input signal for the next iteration ; The first nonlinear detection Input signal for the next iteration The input is fed into the nonlinear estimator for nonlinear detection, and the first nonlinear detection result is obtained. The posterior output of the next iteration ; Determine whether the preset requirements are met; If not satisfied, then the first nonlinear detection will be... Input signal for the next iteration The first nonlinear detection The posterior output of the next iteration The input is fed into the second orthogonal module for orthogonalization processing to obtain the first... The signal after the feature domain of the next iteration is orthogonal ; and the first The signal after the feature domain of the next iteration is orthogonal The input is fed into the cross-domain transformation module to perform inverse cross-domain transformation based on the interleaved frequency division multiplexing matrix, thereby achieving modulation and obtaining the first... Signal before damping in the next iteration ; If satisfied, then the first nonlinear detection... The posterior output of the next iteration As the semantic reconstruction signal output; The receiving end inputs the semantic reconstruction signal to the semantic decoder to complete the target task.

2. The semantic communication method according to claim 1, characterized in that, The first The signal after +1 iterations of orthogonalization of the feature domains satisfies the following formula: ， in, For the first The signal after +1 iterations of orthogonal feature domains Indicates the first Nonlinear detector in the next iteration The first normalized coefficient used for the nonlinear detector One value, Indicates the first Nonlinear estimator at the next iteration The first orthogonal coefficient used for the second orthogonalization module Values.

3. The semantic communication method according to claim 1, characterized in that, The nonlinear estimator is specifically used for: Based on a pre-trained score model, the first nonlinear detection is approximated. Input signal for the next iteration The gradient of the marginal log-likelihood probability; Based on the gradient, the first nonlinear detection Input signal for the next iteration The variance of the nonlinear detection is used for denoising estimation to obtain the first nonlinear detection result. posterior output .

4. The semantic communication method according to claim 3, characterized in that, The first nonlinear detection posterior output Satisfy the following formula: ， in, Indicates the first Nonlinear estimator at the next iteration The nonlinear detection of the first Input signal for the next iteration variance Represents the fractional model The estimated gradient.

5. The semantic communication method according to claim 3, characterized in that, The first nonlinear detection Input signal for the next iteration The variance is updated using the Stan unbiased risk estimation algorithm.

6. The semantic communication method according to claim 5, characterized in that, The first nonlinear detection Input signal for the next iteration The variance is updated using the following formula: ， in, The first nonlinear detection The variance of the input signal in the next iteration For the length of the received signal, Indicates the first Nonlinear estimator at the next iteration express variance Represents the nonlinear estimator pair Find the partial derivative. For element index, To determine the sign of the partial derivative.

7. The semantic communication method according to claim 3, characterized in that, The training methods for the semantic encoder, the semantic decoder, and the score model include: The semantic encoder and the semantic decoder are jointly trained; The fractional model is trained within the framework of variance explosion stochastic differential equations.

8. A semantic communication system based on interleaved frequency division multiplexing, characterized in that, Includes the sending end and the receiving end; The sending end is used for: The source data is encoded based on a semantic encoder to extract the semantic information required by the target task and obtain a semantic signal. The semantic signal is then interleaved with frequency division multiplexing modulation to obtain a time-domain signal. The interleaved frequency division multiplexing modulation matrix has universal characteristics. Transmit the time-domain signal; The receiving end is used for: Based on a cross-domain iterative detector, the received signal is used for iterative updates to recover semantic information and obtain a semantic reconstruction signal; The cross-domain iterative detector includes a time-domain linear detector, a feature-domain nonlinear detector, and a cross-domain transformation module. The time-domain linear detector includes a damping module, a memory matched filter, and a first orthogonal module. The nonlinear detector includes a nonlinear estimator and a second orthogonal module. The cross-domain iterative detector is used to: perform linear detection based on the received signal in the time domain using channel sparsity, perform nonlinear detection based on a fractional model in the feature domain, and realize cross-domain transformation of the signal between the time domain and the feature domain based on the interleaved frequency division multiplexing modulation matrix. The receiving end uses the received signal to perform the first... The process of each iteration update includes: The first Signal before damping in the next iteration The input is sent to the damping module for damping operation to obtain the first... The damped signal after the next iteration ,in, It is a positive integer; The first The damped signal after the next iteration and the received signal The input is fed into the memory matched filter for linear detection, and the first linear detection result is obtained. The posterior output of the next iteration ; The first The damped signal after the next iteration The first linear detection The posterior output of the next iteration The input is fed into the first orthogonal module for orthogonalization processing to obtain the first... The signal after time-domain orthogonality in the next iteration ; The first The signal after time-domain orthogonality in the next iteration The input is fed into the cross-domain transformation module to perform cross-domain transformation based on the interleaved frequency division multiplexing modulation matrix, thereby achieving demodulation and obtaining the first nonlinear detection result. Input signal for the next iteration ; The first nonlinear detection Input signal for the next iteration The input is fed into the nonlinear estimator for nonlinear detection, and the first nonlinear detection result is obtained. The posterior output of the next iteration ; Determine whether the preset requirements are met; If not satisfied, then the first nonlinear detection will be... Input signal for the next iteration The first nonlinear detection The posterior output of the next iteration The input is fed into the second orthogonal module for orthogonalization processing to obtain the first... The signal after the feature domain of the next iteration is orthogonal ; and the first The signal after the feature domain of the next iteration is orthogonal The input is fed into the cross-domain transformation module to perform inverse cross-domain transformation based on the interleaved frequency division multiplexing matrix, thereby achieving modulation and obtaining the first... Signal before damping in the next iteration ; If satisfied, then the first nonlinear detection... The posterior output of the next iteration As the semantic reconstruction signal output; The semantic reconstruction signal is input into the trained semantic decoder to complete the target task.

Images