Channel model driven semantic communication method and system in dynamic scenario

Abstract

The application provides a channel model driven semantic communication method and system in a dynamic scene, the method comprising: establishing and training a communication network model, establishing a communication system according to the trained network model, the communication system comprising a semantic extraction model, a semantic recovery model, a dynamic source channel joint encoder and a dynamic source channel joint encoder; obtaining an original dynamic scene signal, extracting the semantic features of the original dynamic scene signal; the dynamic source channel joint encoder adaptively encodes the semantic features to generate a semantic encoded signal; the semantic encoded signal is transmitted, and in the transmission process, the semantic encoding distortion is converted into a distorted signal; the receiving end receives the distorted signal, and the dynamic source channel joint decoder decodes the distorted signal to recover the semantic features; the semantic recovery module is used for multi-scale fusion reconstruction of the semantic features to generate and output the original semantic signal. The problem of insufficient robustness of semantic communication in a dynamic scene is solved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a channel model-driven semantic communication method and system in dynamic scenarios. Background Technology

[0002] With the development of 6G mobile communication technology, the core of information transmission has shifted from traditional bit-level reliability to semantic-level intelligent interaction. Semantic communication, as a key technology for the integration of communication and artificial intelligence, can directly extract and transmit the semantic connotation of information. It can still ensure the performance of critical tasks in scenarios with extremely low signal-to-noise ratio and limited bandwidth, and has become one of the core supporting technologies of 6G.

[0003] Currently, deep learning-based joint source-channel coding is the mainstream solution for end-to-end optimization of semantic communication, demonstrating excellent performance in multimodal data transmission such as images, videos, text, and voice. Meanwhile, to address the problem of dynamic channel distortion, traditional communication systems generally employ pilot-assisted channel estimation and equalization schemes. These schemes estimate the channel response by inserting pilot signals and perform inverse compensation to counteract amplitude and phase frequency distortions in the channel.

[0004] In dynamic scenarios, video semantic information transmission provides a key basis for autonomous driving environmental perception and multi-vehicle collaborative decision-making. However, the high-speed movement of dynamic terminals and the complex electromagnetic environment in cities can cause significant multipath effects and Doppler frequency shifts, resulting in strong time-varying amplitude and phase frequency distortion characteristics of the channel, which places extremely high demands on the robustness of semantic communication systems. Summary of the Invention

[0005] In view of this, the purpose of this application is to propose a channel model-driven semantic communication method and system in dynamic scenarios, which solves the problem of insufficient robustness of semantic communication in dynamic scenarios.

[0006] To achieve one of the aforementioned objectives, this application provides a channel model-driven semantic communication method for dynamic scenarios, the method comprising:

[0007] Establish a network model; wherein the network model includes a network architecture from the transmitter to the receiver and a dynamic channel model; A network model is trained, and a communication system is established based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model located at the transmitting end, a semantic recovery model located at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine a distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated by the trained dynamic channel model. The original dynamic scene signal is acquired at the transmitting end, and the semantic extraction module is used to extract the semantic features in the original dynamic scene signal. At the transmitting end, the channel response of the dynamic channel model is used as guiding information, and the semantic features are adaptively encoded using the dynamic source-channel joint encoder to generate a semantic coded signal. The semantically encoded signal is transmitted from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and becomes a distorted signal. The receiver receives the distorted signal and uses a dynamic source-channel joint decoder to decode the distorted signal using the channel impulse response provided by the trained dynamic channel model, thereby restoring the semantic features. The semantic recovery module is used to perform multi-scale fusion reconstruction of the semantic features to generate and output the original semantic signal.

[0008] As a further improvement to one embodiment of this application, the distortion signal includes: The distorted signal can be represented by the following formula: ; y = hx + n, n ~ N(0, σ) 2 ); Among them, D e ( ) represents the dynamic source-channel joint encoder, S α ( ) represents the semantic extraction module, and z represents the original dynamic scene signal. Let x be the channel response, h be the semantically encoded signal, n be the true channel response, and σ be the Gaussian noise. 2 Let y be the noise power and y be the distorted signal. The step of using the semantic recovery module to perform multi-scale fusion reconstruction of the semantic features, generating and outputting the original semantic signal, includes: The original semantic signal is generated using the following formula: ; in, S is the original semantic signal. β ( ) represents the semantic recovery module, D d ( ) represents the dynamic source-channel joint encoder. y represents the channel response, and y represents the distortion signal.

[0009] As a further improvement to one embodiment of this application, the establishment of the network model includes a dynamic channel model, comprising: A dynamic channel model with multiple inputs and multiple outputs is constructed. The dynamic channel model includes line-of-sight paths and non-line-of-sight paths characterized by multi-scatter cluster synthesis. The line-of-sight paths and the non-line-of-sight paths are used to reflect disturbances caused by mobility in the dynamic scenario.

[0010] As a further improvement to one embodiment of this application, the disturbance includes a Doppler frequency shift; The establishment of the network model includes: The mobile terminal and temporary base station in the dynamic scene are regarded as the relative receiving end and sending end; A uniform linear array antenna is configured at the receiving end and the transmitting end. The distance vector of the antenna and the Doppler frequency shift caused by the movement of the transmitting end are calculated.

[0011] As a further improvement to one embodiment of this application, the calculation of the antenna's range vector includes: The distance vector is calculated according to the following formula; ; in, Let Φ be the distance vector of the p-th antenna of the transmitting end. T Let θ be the angle of elevation. T For azimuth, δ T The distance between adjacent antennas.

[0012] As a further improvement to one embodiment of this application, the disturbance includes line-of-sight path loss, non-line-of-sight path loss, and small-scale fading. The line-of-sight path loss is calculated using the following formula: ; Among them, PL LoS For the line-of-sight path loss, d 3D Let d be the three-dimensional distance between the transmitter and the receiver. 2D f is the two-dimensional distance between the transmitter and the receiver. c The center frequency of the carrier. h is the distance between the breakpoints. BS h is the antenna height at the base station. UT The antenna height at the terminal; The non-line-of-sight path loss is calculated using the following formula: PL NLoS =max[PL Los 13.54 + 39.08 log 10 (d) 3D )+20log 10 (f) c -0.6 (H)R -1.5); Among them, PL NLoS H represents the non-line-of-sight path loss. R The antenna height of the receiving end; The channel matrix under line-of-sight path loss, non-line-of-sight path loss, and small-scale fading perturbations is calculated using the following formula: ; Where H(t, τ) is the channel matrix, PL[] is the path loss amplitude attenuation factor, and h pq (t, τ) represents the components of the small-scale fading. M R M represents the number of antennas at the receiving end. τ The number of antennas at the transmitting end.

[0013] As a further improvement to one embodiment of this application, before decoding the distorted signal using the channel impulse response provided by the trained dynamic channel model, the following steps are included: Calculate the distance vectors of multiple propagation paths between the receiver and the transmitter; The channel impulse response between an antenna at the receiving end and an antenna at the transmitting end is calculated based on multiple distance vectors.

[0014] As a further improvement to one embodiment of this application, the calculation of the distance vector of multiple propagation paths between the receiving end and the transmitting end includes: The propagation path is divided into the path from the transmitter to the first scatterer, the virtual link between the multiple scatterers, and the path from the last scatterer to the receiver. Calculate the time-varying distance vector from the transmitter to the first scatterer based on the moving speed of the transmitter; Calculate the stationary distance vector of the virtual link between the multiple scatterers, and the stationary distance vector from the last scatterer to the receiver.

[0015] As a further improvement to one embodiment of this application, the trained network model includes: The communication system is optimized using the following formula as the optimization objective: ; Among them, P dynamic This refers to the joint distribution of the source signal and channel response in the dynamic scenario described. D is a measure of semantic layer distortion. e For the dynamic source-channel joint encoder, D d For the dynamic source-channel joint decoder, z is the original dynamically refined signal, and h follows the sampling of this dynamic joint distribution. The original semantic signal is output; The optimized communication system is used to dynamically compensate for channel distortion and generate the original semantic signal.

[0016] Based on the same inventive concept, this application also provides a channel model-driven semantic communication system for dynamic scenarios, including: A module is established to build a network model; wherein, the network model includes a network architecture from the transmitter to the receiver and a dynamic channel model; A training module is used to train a network model and establish a communication system based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model located at the transmitting end, a semantic recovery model located at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine a distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated from the trained dynamic channel model. An acquisition module is used to acquire the original dynamic scene signal at the sending end and extract semantic features from the original dynamic scene signal using the semantic extraction module. The encoding module is used at the transmitting end to take the channel response of the dynamic channel model as guiding information, and use the dynamic source-channel joint encoder to adaptively encode the semantic features to generate a semantic encoded signal. A transmission module is used to transmit the semantically encoded signal from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and becomes a distorted signal. The decoding module is used to receive the distorted signal using the receiving end, and decode the distorted signal using the channel impulse response provided by the trained dynamic channel model through the dynamic source-channel joint decoder to restore the semantic features; The output module is used to perform multi-scale fusion reconstruction of the semantic features using the semantic recovery module to generate and output the original semantic signal.

[0017] Compared to existing technologies, the technical advantages of this invention are as follows: The dynamic channel model constructed in this application accurately characterizes the channel distortion mechanism. Combined with a distance-priority adaptive network module, it achieves adaptive compensation for amplitude-frequency and phase-frequency distortion. In core evaluation metrics, it significantly outperforms existing source-channel joint coding schemes based on static channel assumptions. Furthermore, under low signal-to-noise ratio conditions, its performance surpasses pilot-assisted channel estimation and equalization schemes based on ideal channel estimation, effectively combating dynamic time-varying channel distortion and significantly improving semantic reconstruction quality. Even in extreme scenarios such as high mobility and long distances, it maintains stable semantic reconstruction quality, while existing static channel schemes experience a sharp performance decline in such scenarios, verifying the strong adaptability and generalization ability of this application to complex dynamic scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the implementation methods or related technologies will be briefly introduced below. Obviously, the drawings described below are only the implementation methods of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a channel model-driven semantic communication method in a dynamic scenario, provided as an embodiment of this application; Figure 2 A dynamic scene provided for one embodiment of this application; Figure 3 The dynamic scene spatial relationship provided in one embodiment of this application; Figure 4 A schematic diagram illustrating the variation of channel impulse response with distance in a dynamic channel model provided in one embodiment of this application; Figure 5 A schematic diagram of the network architecture of a communication system provided in one embodiment of this application; Figure 6 A schematic diagram of the structure of a distance adaptive network provided in one embodiment of this application; Figure 7 A schematic diagram of a channel model-driven semantic communication system in a dynamic scenario is provided as another embodiment of this application; Figure 8 This is a schematic diagram of the hardware structure of an electronic device provided for another embodiment of this application. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by those skilled in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0022] Existing technologies have significant shortcomings in dynamic channel scenarios, failing to meet the requirements of low latency and high robustness in semantic communication in dynamic scenarios. Specific problems are as follows: First, the channel assumptions are out of touch with real-world scenarios. Mainstream deep learning-based source-channel joint coding schemes are all based on ideal time-invariant channel design and verification, without considering the time-varying distortion of composite amplitude and phase frequencies in dynamic scenarios. In real-world dynamic scenarios, the source-channel mapping relationship learned by the semantic codec fails, leading to semantic feature distortion and a sharp decline in system performance.

[0023] Secondly, pilot-assisted channel estimation and equalization schemes have poor adaptability. Traditional pilot-assisted channel estimation and equalization schemes require the insertion of pilot signals, which not only consumes bandwidth resources and contradicts the original design intention of high compression ratio in semantic communication, but also requires increasing pilot density to ensure estimation accuracy in dynamic, high-speed, time-varying channels, further increasing communication overhead and computational complexity. The introduced estimation delay cannot meet the ultra-reliable low latency requirements of 6G.

[0024] Third, the optimization objectives are mismatched. The optimization objective of pilot-assisted channel estimation and equalization is to minimize the bit / symbol error rate, while the core objective of semantic communication is to ensure the performance of semantic tasks. The optimization directions of the two are inconsistent and cannot effectively adapt to the needs of semantic communication.

[0025] Fourth, pilot-dependent adaptive semantic communication schemes consume bandwidth resources. Current channel adaptive semantic communication schemes all rely on pilots to obtain channel state information, resulting in additional bandwidth consumption and non-real-time feedback. Furthermore, they can only adapt to single channel characteristics and cannot handle complex composite channel distortions in dynamic scenarios.

[0026] To address the aforementioned problems, this application provides a channel model-driven semantic communication method for dynamic scenarios, such as... Figure 1 As shown, the method includes the following steps: Step S100: Establish a network model; wherein the network model includes the network architecture from the transmitter to the receiver and a dynamic channel model.

[0027] Specifically, a network model is first established that departs from the assumption of an ideal time-invariant channel, incorporating dynamic channel characteristics as adversarial perturbations into the end-to-end optimization framework. The dynamic semantic communication system based on this channel model enables pilot-free adaptive semantic communication. This fundamentally breaks through the traditional communication model of transmission before computation, constructing a dynamic semantic communication system architecture that computes first and then transmits.

[0028] In one possible implementation of this application, step S100, establishing the dynamic channel model in the network model, includes: Step S110: Construct a dynamic channel model with multiple inputs and multiple outputs. The dynamic channel model includes line-of-sight paths and non-line-of-sight paths characterized by multi-scatter cluster synthesis. The line-of-sight paths and non-line-of-sight paths are used to reflect disturbances caused by mobility in dynamic scenarios.

[0029] Specifically, to accurately characterize dynamic channel properties and address the complex electromagnetic environment of dynamic scenarios, a 3D non-stationary multiple-input multiple-output (MIMO) dynamic channel model based on the Geometry-Based Stochastic Channel Model (GBSM) is constructed. This model comprehensively represents both line-of-sight (LoS) and non-line-of-sight (NLoS) paths through multiple scattering clusters, effectively reflecting disturbances caused by mobility, such as the Doppler effect, path loss, shadowing fading, and small-scale fading.

[0030] In one possible implementation of this application, the perturbation includes a Doppler frequency shift, and step S100 further includes: Step S121: The mobile terminal and temporary base station in the dynamic scene are regarded as the relative receiver and transmitter.

[0031] Step S122: Configure uniform linear array antennas at the receiving end and the transmitting end, calculate the distance vector of the antennas, and the Doppler frequency shift caused by the movement of the transmitting end.

[0032] Specifically, such as Figure 2 The diagram illustrates the geometric model of a dynamic scene, clearly defining the communication transceivers (mobile terminal Tx, temporary base station Rx), the 3D spatial distribution of the scattering clusters, and the signal propagation paths for line-of-sight, single-hop scattering, and double-hop scattering. This dynamic scene includes a mobile terminal Tx and a temporary base station Rx, both located in a three-dimensional coordinate system. shaft and On the axis. The model uses a cluster delay line (CDL) structure to characterize the broadband channel and frequency selectivity. Both the transmitter and receiver are equipped with uniform linear array (ULA) antennas, with M antennas respectively. T and M R The spacing between adjacent antennas is δ T and δ R .

[0033] In one possible implementation of this application, calculating the antenna's range vector includes: Calculate the distance vector using the following formula; ; in, Let Φ be the range vector of the p-th antenna at the transmitting end. T Let θ be the angle of elevation. T For azimuth, δ T This represents the spacing between adjacent antennas.

[0034] like Figure 3 The diagram illustrates the scattering cluster structure and signal propagation angle / distance parameters of the refined dynamic channel model, demonstrating the correlation between the signal transmission path and channel parameters for single-hop and double-hop scattering.

[0035] The first antenna element of the transmitter is located at coordinate P. T At point P, the first antenna element of the receiver is located at coordinate P. R At this point, the distance vector pointing to the p-th (p=1,2,…,P) antenna at the transmitting end. With elevation angle Φ T With azimuth θ T The initial distance can be expressed as d0 = P. R -P T .

[0036] It should be noted that the distance vector at the receiving end... This can be calculated using the same method. In this scenario, the transmitter travels at a speed of v. T (t) The receiver and the scatterer are assumed to be stationary. Therefore, the Doppler frequency shift is mainly caused by the movement of the mobile terminal.

[0037] In one possible implementation of this application, calculating the Doppler frequency shift caused by the movement of the transmitting end includes: The Doppler frequency shift of the line-of-sight path is calculated using the following formula: ; in, For the Doppler frequency shift of the line-of-sight path, v is the distance vector of the line-of-sight path.T (t) represents the moving speed of the transmitter, and λ represents the signal wavelength.

[0038] In one possible implementation of this application, the disturbance includes line-of-sight path loss, non-line-of-sight path loss, and small-scale fading. Step S100 further includes: Step S131: Calculate the line-of-sight path loss according to the following formula: ; Among them, PL LoS For line-of-sight path loss, d 3D Let d be the three-dimensional distance between the transmitter and receiver. 2D Let f be the two-dimensional distance between the transmitter and receiver. c The center frequency of the carrier. h is the distance between the breakpoints. BS h is the antenna height at the base station. UT This refers to the antenna height at the terminal.

[0039] Step S132, calculate the non-line-of-sight path loss according to the following formula: PL NLoS =max[PL Los 13.54 + 39.08 log 10 (d) 3D )+20log 10 (f) c -0.6 (H) R -1.5); Among them, PL NLoS For non-line-of-sight path loss, H R This represents the antenna height at the receiving end.

[0040] It should be noted that for large-scale fading, assuming the signal propagation path is in a densely populated urban environment with buildings and vehicles, the impact of obstacles cannot be ignored. Unlike free-space propagation, signals experience significant attenuation during propagation due to reflection and diffraction. In engineering, empirical formulas and models are typically used to characterize such environments. The 3GPP Urban Macro (UMa) model is one of the standard empirical models in the field of wireless mobile communication. This model is mainly applicable to approximately flat areas such as urban streets and suburbs where the transmitter height is lower than surrounding buildings. This is based on the 3GPP UMa scenario path loss model. Furthermore, the loss expression for non-line-of-sight paths needs to consider the impact of obstacle occlusion.

[0041] The channel matrix under line-of-sight path loss, non-line-of-sight path loss, and small-scale fading perturbations is calculated using the following formula: ; Where H(t, τ) is the channel matrix, PL[] is the path loss amplitude attenuation factor, and h pq (t, τ) represents the components of small-scale fading. M R M represents the number of antennas at the receiving end. τ This represents the number of antennas at the transmitting end.

[0042] It is important to note that for small-scale fading, the distribution characteristics of scattering clusters need to be emphasized. Scattering clusters in a channel refer to scatterers aggregated in the environment. Therefore, a geometrically random channel model is needed to capture the spatiotemporal characteristics of the channel. Unlike static scenarios, the topology of dynamic scenarios exhibits time-varying characteristics as the transmitter moves. The model divides the scatterers into different clusters, including single-bounce (SB) clusters near the transmitter and receiver, and double-bounce (DB) clusters representing far-field scattering. Each scattering cluster forms an independent propagation path with a distinguishable time delay.

[0043] In other possible implementations, the parameters of the dynamic channel model, such as the number of scattering clusters, the number of scatterers, the Rice K-factor, and the carrier frequency, can be adjusted according to different dynamic scenarios without changing the core design of 3D non-stationary MIMO and geometric stochastic modeling.

[0044] Step S200: Train the network model and establish a communication system based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model at the transmitting end, a semantic recovery model at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine the distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated by the trained dynamic channel model.

[0045] In one possible implementation of this application, step S200 includes: Step S210: Optimize the communication system using the following formula as the optimization objective: ; Among them, P dynamic This represents the joint distribution of source signal and channel response in dynamic scenarios. D is a measure of semantic layer distortion. e For a dynamic source-channel joint encoder, D d For a dynamic source-channel joint decoder, z is the original dynamically refined signal, and h follows the sampling of this dynamic joint distribution. This is the original semantic signal output. Step S220: Dynamically compensate for channel distortion using the optimized communication system to generate the original semantic signal.

[0046] Specifically, such as Figure 4 As shown, four sub-figures illustrate: (a) channel response amplitude at a terminal speed of 10 m / s; (b) channel response amplitude at a terminal speed of 100 m / s; (c) 3D channel response at a terminal speed of 10 m / s; and (d) 3D channel response at a terminal speed of 100 m / s. These figures visually demonstrate the changes in amplitude-frequency and phase-frequency distortion characteristics of the channel as communication distance and terminal speed increase. With increased communication distance and terminal speed, the channel impulse response exhibits significant time dispersion and Doppler spread, leading to a sharp decline in the performance of models trained using traditional static channels. The core issue boils down to how to adaptively compensate for the composite amplitude-frequency and phase-frequency distortion of time-varying channels without relying on pilots and with low computational complexity constraints, ensuring the fidelity and robustness of semantic transmission. Dynamic semantic communication systems based on channel models utilize available channel model information. This allows for dynamic compensation of channel distortion, thereby maintaining high semantic fidelity in time-varying environments. Therefore, the entire communication system is optimized.

[0047] To address the complex distortion characteristics of dynamic channels, this application designs an adaptive communication system based on channel model priors, a prior-driven network, and a multi-stage training method to achieve low-overhead, highly robust semantic transmission. The complex semantic distortion caused by the channel is strongly correlated with the distance between the transmitter and receiver. This system design abandons the reliance on ideal channel state information, instead using readily available initial and current distances as prior knowledge to drive the encoding and decoding process for efficient adaptation and adaptive countermeasure against channel distortion. The complex distortion caused by the channel can be decoupled into two parts: a deterministic component strongly correlated with distance and a random component related to the scattering environment. The dynamic channel adaptive semantic communication architecture aims to compensate for the deterministic component through channel model priors, while simultaneously adapting to the random component using the generalization ability of deep learning models.

[0048] Step S300: Obtain the original dynamic scene signal at the sending end, and use the semantic extraction module to extract the semantic features in the original dynamic scene signal.

[0049] In step S400, the channel response of the dynamic channel model is used as guiding information at the transmitting end, and the semantic features are adaptively encoded using the dynamic source-channel joint encoder to generate a semantic coded signal.

[0050] In step S500, the semantically encoded signal is transmitted from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and converted into a distorted signal.

[0051] In one possible implementation of this application, step S500 converts the signal into a distorted signal, including: Step S510, the distortion signal is represented by the following formula: ; y = hx + n, n ~ N(0, σ) 2 ); Among them, D e ( ) represents the dynamic source-channel joint encoder, S α ( ) represents the semantic extraction module, and z represents the original dynamic scene signal. Let x be the channel response, h be the semantically encoded signal, n be the true channel response, and σ be the Gaussian noise. 2 y represents noise power and y represents the distortion signal.

[0052] Specifically, at the transmitting end, through the semantic extraction module S α ( ) Extract highly compressed semantic features from the original dynamic scene signal z. Then, the channel response obtained from the dynamic channel model... As guiding information, it is transmitted through the dynamic source-channel joint coding module D. e ( ) The extracted multimodal semantic features are adaptively encoded to obtain the encoded signal x, which can dynamically adapt to time-varying channel conditions.

[0053] It should also be noted that the signal-to-noise ratio during signal transmission is expressed as: ; Where SNR is the signal-to-noise ratio, S is the signal power, and σ is the signal strength. 2 This represents noise power.

[0054] Step S600: The receiver receives the distorted signal and the dynamic source-channel joint decoder uses the channel impulse response provided by the dynamic channel model to decode the distorted signal and restore the semantic features.

[0055] Specifically, at the receiving end, the dynamic source-channel joint decoder D... d ( ) Utilize the Channel Impulse Response (CIR) provided by the dynamic channel model. Semantic features are directly recovered from the distorted signal y, and finally processed by the semantic recovery module S. β ( ) Reconstruct the original semantic signal , is represented as: .

[0056] In one possible implementation of this application, before step S600 decodes the distorted signal using the channel impulse response provided by the dynamic channel model, the following steps are included: Step S601: Calculate the distance vectors of multiple propagation paths between the receiver and the transmitter.

[0057] Specifically, considering that only the transmitting end is in motion, the distance vector of the line-of-sight path can be described as: ; in, v is the distance vector of the line-of-sight path. T (t) represents the moving speed of the transmitter, P T Let P be the coordinates of the antenna at the transmitting end. R Here are the coordinates of the antenna at the receiving end. Let be the distance vector of the p-th antenna at the receiving end. Let be the distance vector of the p-th antenna at the transmitting end.

[0058] Step S602: Calculate the channel impulse response between an antenna at the receiving end and an antenna at the transmitting end based on multiple distance vectors.

[0059] Specifically, it is calculated according to the following formula: ; ; ; ; in, Let KR be the channel matrix between antenna p and antenna q, and KR be the Rice K-factor, representing the proportion of the line-of-sight (LoS) component. The impulse response of the p-th transmitting and q-th receiving antennas in the line-of-sight channel. The percentage of single-hop scattering power. The single-hop channel impulse response between the p-th transmitting and q-th receiving antennas. This represents the percentage of power used for double-hop scattering. For Loss path delay, This represents the number of single-hop scattering clusters. This represents the number of double-hop scattering clusters. Let be the instantaneous power of the l-th single-hop scattering path at time t. The complex exponential factor corresponding to the random phase shift introduced into the m-th scatterer in the l-th single-hop scattering cluster. The impulse response of the p-th transmit and q-th receive antennas in a double-hop channel. For the m-th scatterer within the l-th cluster in a double-hop path, the random phase shift introduced during the first reflection... Let be the three-dimensional position vector from the p-th transmitting antenna to the m-th scatterer within the l-th cluster at time t. Let be the virtual equivalent scattering point position vector corresponding to the m-th double-hop path within the l-th cluster at time t. Let be the three-dimensional position vector from the m-th scatterer in the l-th cluster to the q-th receiving antenna at time t.

[0060] The power of the scattering cluster satisfies η SB+η DB =1. According to the 3GPP TR 38.901 protocol specification, the power of a single reflection cluster is... With secondary reflection cluster power The specific values ​​can be calculated using the following formula: ; Among them, Z l Let Z be a random variable that follows a normal distribution. l ~N(0,ζ) 2 ), ζ is the standard deviation. and τ l (t) represents the time delay spread and time delay of the l-th scattering cluster, respectively, r τ This is the time delay scaling factor. To ensure the total power is normalized, the calculation results need to be normalized. .

[0061] The definitions of the above parameter symbols are shown in Table 1.

[0062] Table 1 - Parameter Symbols and Definitions

[0063] In one possible implementation of this application, step S601 includes: Step S6011: The propagation path is divided into the path from the transmitter to the first scatterer, the virtual link between multiple scatterers, and the path from the last scatterer to the receiver.

[0064] Step S6012: Calculate the time-varying distance vector from the transmitter to the first scatterer as the transmitter moves.

[0065] Step S6013: Calculate the stationary distance vector of the virtual link between multiple scatterers, and the stationary distance vector from the last scatterer to the receiver.

[0066] Specifically, the range vector of the secondary reflection path consists of three parts: the path from the transmitter to the first scatterer, the virtual link between the scatterers, and the path from the second scatterer to the receiver. Since both the scatterers and the receiver remain stationary, the range vector connected only to the transmitter will exhibit time-varying characteristics with the transmitter's velocity.

[0067] Calculate the time-varying distance vector and the stationary distance vector using the following formulas: ; ; ; ; ; .

[0068] The meanings of the parameters are shown in Table 1.

[0069] Distance vector of a single reflection path and The same method can be used to calculate the vector at the transmitting end, which includes a velocity integral term, while the vector at the receiving end remains stationary.

[0070] It should be added that the Doppler frequency shift is caused by the relative motion between the transmitter and the scatterer or receiver. Since both the receiver and the scatterer are stationary, the Doppler frequency shift of the secondary reflection path can be calculated using the following formula: ; in, For the Doppler frequency shift of the secondary reflection path, || ‖ represents the Euclidean norm. λ represents the inner product, and λ represents the signal wavelength.

[0071] Doppler frequency shift of a single reflection path It has the same expression form as the secondary reflection path, and its value is determined only by the transmitter velocity vector and the signal departure angle.

[0072] Meanwhile, in the cluster delay line structure, both the distribution range of the scattering clusters and the propagation direction of the scattered rays are restricted. Therefore, for the arrival and departure angles of the scattering clusters, this model assumes that they follow a truncated Gaussian distribution, with the probability density function of which is: ; in, ( ) and Φ( ) are the probability density function (PDF) and cumulative distribution function (CDF) of the Gaussian distribution, respectively. With σ x Let x be the mean and variance of a Gaussian distributed random variable truncated from x. low With x up These represent the lower and upper bounds of the random variable x, respectively. By setting the parameters appropriately, the distribution range of the scattering cluster angle can be restricted to a Gaussian distribution within a specific interval.

[0073] Since buildings are relatively evenly distributed within urban areas, this model assumes that the distance distribution of the scattering clusters follows a uniform distribution, and its probability density function is: ; Where a is the lower bound of the uniform distribution and b is the upper bound of the uniform distribution.

[0074] To gain a deeper understanding of how dynamic channels affect semantic communication systems, this application analyzes channel impulse response. The mechanism by which semantic communication is affected lays the theoretical foundation for the design of subsequent channel adaptive algorithms.

[0075] The amplitude attenuation in the channel response is mainly determined by path loss.

[0076] ; .

[0077] These two formulas show that path loss is closely related to 3D propagation distance, and can be expressed as: .

[0078] This attenuation effect leads to semantic features The total energy decreases with increasing propagation distance. Terminal movement is the main source of phase-frequency distortion. As the foregoing analysis shows, phase-frequency characteristics are mainly affected by the exponential component in the channel impulse response. The phase expression for the line-of-sight path is: ; in, The phase of the line-of-sight path, For the Doppler frequency shift of the line-of-sight path, This is the distance vector of the line-of-sight path.

[0079] The non-line-of-sight path phase is determined by the multipath superposition effect of the scatterer, and can be expressed as: ; in, For non-line-of-sight paths, the phase is... and i represents the path The time-varying propagation distance vector, The Doppler frequency shift is the path.

[0080] Step S700: Use the semantic recovery module to perform multi-scale fusion reconstruction of semantic features, generate and output the original semantic signal.

[0081] In one possible implementation of this application, step S700 includes: Step S701: Generate the original semantic signal using the following formula: ; in, S is the original semantic signal. β ( ) represents the semantic recovery module, Dd ( ) represents the dynamic source-channel joint encoder. Let y be the channel response and y be the distortion signal.

[0082] Another implementation of this application discloses a channel model-driven semantic communication system in a dynamic scenario, such as... Figure 5 As shown, Figure 5 This paper demonstrates the end-to-end network architecture of a channel model-driven semantic communication system, including a semantic extraction module and a dynamic source-channel joint encoder combined with a distance adaptive network at the transmitter end, a dynamic source-channel joint decoder combined with a distance adaptive network, a semantic recovery module, and a dynamic channel module at the receiver end. The paper clarifies the input, output, and data transmission flow of each module and marks the input position of the distance prior.

[0083] A symmetrical end-to-end architecture is adopted, driven by prior information from the channel model, integrating semantic representation and adaptive channel coding capabilities. The residual block (ResBlock) fuses the input features with the output features processed by two rounds of Rectified Linear Unit (ReLU) + convolutional layers. The overall architecture of dynamic channel adaptive semantic communication adopts a symmetrical structure and can be divided into three modules: a semantic extraction and recovery module, a dynamic joint source-channel coding and decoding module (dynamic JSCC coding and decoding module), and a wireless channel module.

[0084] The semantic extraction and reconstruction module is responsible for end-to-end representation learning of semantic information. It consists of two symmetrical networks, both employing ResNet-based convolutional neural networks (CNNs). The kernel sizes of the semantic extractor are designed in descending order, enabling the extraction of multi-level, highly compressed semantic features from the original RGB space, minimizing semantic information loss and balancing semantic fidelity and visual quality. Its output provides semantically rich feature representations for subsequent joint source-channel coding. The semantic reconstructor then reconstructs the image in RGB space from the decoded semantic features. Through multi-scale feature fusion techniques, this network ensures optimal performance in both semantic fidelity and visual quality of the reconstructed signal.

[0085] In other possible implementations, the network architecture of the semantic module can replace CNN with networks such as Swin Transformer and MobileViT to adapt to the low computing power hardware environment of in-vehicle terminals without changing the core design of multi-scale semantic feature extraction and fusion.

[0086] The dynamic joint source-channel coding and decoding module is the core of realizing channel adaptation capability. The dynamic joint source-channel coding encoder introduces a channel model adaptation network based on the traditional deep joint source-channel coding encoder, using distance priors d0 and d... t As conditional inputs, semantic features are dynamically encoded into transmission symbols. This process ensures that the core semantic information of the transmitted signal is effectively preserved even after channel distortion. The dynamic joint source-channel codec and decoder employs a symmetric design and also embeds a channel model adaptive network. t Predicting channel distortion enables accurate recovery of semantic features.

[0087] In other possible implementations, the distance prior is obtained by considering the initial distance d0 and the current distance d between the transmitting and receiving ends. t Distance information can be obtained through various methods such as Global Positioning System (GPS), BeiDou positioning, and unmanned vehicle ranging protocols. As long as accurate distance information can be provided in real time, it can be used as a priori input for distance adaptive networks.

[0088] In the wireless channel module, transmitted symbols are processed through a dynamic channel model tailored to the characteristics of high-orbit satellite scenarios. This model accurately simulates time-varying effects such as multipath fading, Doppler shift, and path loss, ultimately outputting a received signal with composite distortion. Through this architecture, the JSCC-based dynamic channel adaptive semantic communication architecture achieves end-to-end optimization of semantic feature extraction, adaptive coding, channel transmission, and distortion compensation.

[0089] The adaptive network based on channel model priors employs a hybrid architecture that integrates a lightweight multilayer perceptron (MLP) with an attention mechanism. Leveraging the MLP's ability to learn complex nonlinear mappings, this network constructs distance priors d0 and d... t A correlation model between channel distortion and adaptive compensation for amplitude attenuation and phase shift caused by channel distortion is established.

[0090] In other possible implementations, the fully connected layer dimension of the MLP structure of the distance adaptive network module can be adjusted according to the actual dynamic scenario, and the ReLU activation function can be replaced with lightweight activation functions such as GELU and Swish, without changing the core design of distance prior-driven amplitude / phase compensation.

[0091] like Figure 6 As shown, Figure 6The structure of the distance adaptive network is shown, including the complex symbol mapping of the input features, the amplitude compensation branch, the phase compensation branch, and the final compensated complex symbol output, clarifying the implementation process of amplitude and phase compensation respectively.

[0092] The input features are mapped to complex numbers in pairs of feature elements, and then the initial amplitude value and phase angle are calculated according to the following formula for split-path correction: ; ; Where A is the initial amplitude value, a and b are two characteristic elements, and θ is the phase angle.

[0093] In the amplitude compensation phase, firstly, average pooling is performed on the initial amplitude to extract global statistical features, and then it is compared with the distance prior. and The features are then fused. The fused features undergo a non-linear mapping via a multilayer perceptron to generate an adaptive weight vector. Multiplying the weight vector by the initial magnitude yields the compensated magnitude, expressed as: ; in, This is the range after compensation.

[0094] In the phase compensation stage, the distance prior is first mapped to a standard interval through a normalization layer. Then, a multilayer perceptron is used to perform feature transformation on the normalized distance prior to obtain the normalized rotated phase. After performing inverse normalization to obtain the true rotation phase, it is compared with the initial phase. Subtraction is used to cancel out the phase shift, and the compensated phase is finally obtained, expressed as: ; in, The compensated phase, This is the phase shift.

[0095] The compensated complex number symbol can be represented as: ; The corresponding compensated two-bit feature elements are: , .

[0096] To ensure that the channel model-driven adaptive semantic communication system can stably and efficiently learn the adaptive encoding and decoding mapping relationship under dynamic channels, a three-stage progressive training strategy was designed.

[0097] This strategy decomposes the complex end-to-end optimization problem into three sub-tasks, effectively avoiding the problem of network parameters getting trapped in local optima due to simultaneously learning semantic representations and channel adversarial mechanisms, thereby improving the model's convergence speed and final performance. The training loss functions corresponding to the three stages are as follows: ; ; ; in, The mean square error loss between the final output signal and the original signal is expressed as: ; in, Let m be the number of pixels in the vertical direction of the image, n be the number of pixels in the horizontal direction of the image, u be the vertical coordinate index of the pixel, and v be the horizontal coordinate index of the pixel.

[0098] Similarly, This represents the loss between the initial phase and the compensated phase, and its expression is consistent with the mean square error loss described above.

[0099] The first stage involves pre-training the semantic representation, freezing the dynamic modules, and training only the semantic extractor and restorer, using mean squared error (MSE) loss. This ensures effective extraction and reconstruction of semantic features. The second stage involves training adaptive channel coding, freezing the semantic module weights, and integrating the dynamic channel model into the system, inputting the distance priors d0 and d... t Training the dynamic encoder-decoder using phase The third stage involves end-to-end joint optimization, unlocking all module weights and employing hybrid loss. By balancing the weights of the two hyperparameters λ1 and λ2, global optimization of semantic representation and channel adaptation is achieved.

[0100] This application proposes a channel model-driven semantic communication method and system for dynamic scenarios. One key feature is the design of a distance-adaptive network that integrates distance priors. Addressing the strong correlation between dynamic channel distortion and the distance between the transmitter and receiver, a lightweight distance-adaptive network is designed. This network integrates the initial and current distances as prior information and employs a hybrid architecture of multilayer perceptron and attention mechanism to compensate for the amplitude and phase of semantic features separately. By learning the nonlinear mapping relationship between distance and channel amplitude-frequency and phase-frequency distortion through the multilayer perceptron, adaptive weighted compensation for amplitude attenuation and adaptive rotational compensation for phase shift are achieved, effectively combating the compound distortion of the channel. Pilot overhead is eliminated, reducing computational and communication complexity. No pilot signal insertion is required, allowing all bandwidth resources to be used for semantic feature transmission, improving spectral efficiency. Furthermore, the complex operations of traditional pilot-assisted channel estimation and equalization are abandoned. Adaptive compensation is achieved solely through the lightweight distance-adaptive network, with only a slight increase in model parameters. The receiver only needs one forward propagation to complete semantic reconstruction, significantly reducing computational complexity and transmission delay, meeting the low-complexity requirements of dynamic scenarios.

[0101] In addition, a dynamic joint source-channel coding module was designed. A distance-prior adaptive network was integrated into the source-channel joint coding and decoding process to construct the dynamic source-channel joint coding module: the transmitter's dynamic source-channel joint encoder, using distance prior as a condition, adaptively encodes the semantic features output by the semantic extraction module, generating robust transmission symbols insensitive to channel distortion; the receiver's dynamic source-channel joint decoder, also incorporating distance prior, directly adaptively decodes the received signal with time-varying channel distortion, eliminating the need for channel estimation and equalization, thus achieving semantic communication where computation precedes transmission. This effectively combats dynamic time-varying channel distortion, significantly improves semantic reconstruction quality, and outperforms existing source-channel joint coding schemes based on static channel assumptions in core evaluation metrics. Furthermore, its performance surpasses pilot-assisted channel estimation and equalization schemes with ideal channel estimation at low signal-to-noise ratios.

[0102] Furthermore, to ensure stable convergence and high performance of the model under dynamic channels, a three-stage progressive training strategy is designed. The first stage trains the semantic extraction and reconstruction modules separately to ensure effective extraction and reconstruction of semantic features. The second stage freezes the weights of the semantic modules and trains the distance-adaptive coding and decoding capabilities of the dynamic source-channel joint coding module. The third stage jointly trains all modules, aiming to minimize semantic distortion at the semantic level, achieving end-to-end adaptive semantic transmission optimization. This approach adapts to extreme dynamic scenarios. The technical solution of this application maintains stable semantic reconstruction quality even under extreme scenarios such as high mobility and long distances, while the performance of existing static channel solutions deteriorates sharply in such scenarios, verifying the strong adaptability and generalization ability of this application to complex dynamic scenarios.

[0103] This application incorporates the dynamic characteristics of the channel as a means of mitigating disturbances into the joint coding design of the source and channel. It prioritizes minimizing semantic-level distortion as its optimization objective, rather than the traditional bit / symbol error rate, perfectly meeting the core requirements of semantic communication. Furthermore, the solution is based on end-to-end deep learning, making it easy to integrate with existing dynamic communication hardware, thus providing a feasible solution for the practical deployment of semantic communication in 6G dynamic scenarios.

[0104] Another embodiment of this application discloses a channel model-driven semantic communication system in dynamic scenarios, such as... Figure 7 As shown, it includes: The module is used to build the network model; the network model includes the network architecture from the transmitter to the receiver and the dynamic channel model. The training module is used to train the network model and establish a communication system based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model at the transmitting end, a semantic recovery model at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine a distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated by the trained dynamic channel model. The acquisition module is used to acquire the original dynamic scene signal at the sending end and extract the semantic features from the original dynamic scene signal using the semantic extraction module. The encoding module is used at the transmitting end to take the channel response of the dynamic channel model as the guiding information, and use the dynamic source-channel joint encoder to adaptively encode the semantic features to generate a semantic coded signal. The transmission module is used to transmit semantically encoded signals from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and becomes a distorted signal. The decoding module is used to receive distorted signals at the receiver and decode the distorted signals using the channel impulse response provided by the trained dynamic channel model through a dynamic source-channel joint decoder, thereby restoring the semantic features. The output module is used to perform multi-scale fusion reconstruction of semantic features using the semantic recovery module, and generate and output the original semantic signal.

[0105] Figure 8 This diagram illustrates a more specific hardware structure of an electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0106] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0107] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0108] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0109] The communication interface 1040 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, radio (shortwave / ultra-shortwave) communication, satellite communication, data link communication, etc.).

[0110] Bus 1050 includes pathways for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0111] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments described in this specification, and need not include all the components shown in the figures.

[0112] The electronic device described above is used to implement the channel model-driven semantic communication method in the corresponding dynamic scenario in any of the foregoing embodiments, and has the beneficial effects of the corresponding method implementation, which will not be elaborated here.

[0113] Based on the same inventive concept, corresponding to any of the above-described embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the channel model-driven semantic communication method in a dynamic scenario as described in any of the above embodiments.

[0114] The computer-readable medium in this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0115] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the channel model-driven semantic communication method in dynamic scenarios as described in any of the above embodiments, and have the beneficial effects of the corresponding method implementation, which will not be repeated here.

[0116] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; this manner of description is merely for clarity, and those skilled in the art should consider the specification as a whole. Within the framework of this application, the above embodiments or the technical features of different embodiments can also be appropriately combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0117] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be entirely within the understanding of those skilled in the art). While specific details (e.g., circuits) are set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0118] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0119] The embodiments described herein are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and principles of the embodiments described herein should be included within the protection scope of this application.

Claims

1. A channel model-driven semantic communication method for dynamic scenarios, characterized in that, The method includes: Establish a network model; wherein the network model includes a network architecture from the transmitter to the receiver and a dynamic channel model; A network model is trained, and a communication system is established based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model located at the transmitting end, a semantic recovery model located at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine a distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated by the trained dynamic channel model. The original dynamic scene signal is acquired at the transmitting end, and the semantic extraction module is used to extract the semantic features in the original dynamic scene signal. At the transmitting end, the channel response of the dynamic channel model is used as guiding information, and the semantic features are adaptively encoded using the dynamic source-channel joint encoder to generate a semantic coded signal. The semantically encoded signal is transmitted from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and becomes a distorted signal. The receiver receives the distorted signal and uses a dynamic source-channel joint decoder to decode the distorted signal using the channel impulse response provided by the trained dynamic channel model, thereby restoring the semantic features. The semantic recovery module is used to perform multi-scale fusion reconstruction of the semantic features to generate and output the original semantic signal.

2. The channel model-driven semantic communication method in dynamic scenarios according to claim 1, characterized in that, The distorted signal includes: The distorted signal can be represented by the following formula: ； y=hx+n, n~N(0, σ) 2 ); Among them, D e ( ) represents the dynamic source-channel joint encoder, S α ( ) represents the semantic extraction module, and z represents the original dynamic scene signal. Let x be the channel response, h be the semantically encoded signal, n be the true channel response, and σ be the Gaussian noise. 2 Let y be the noise power and y be the distorted signal. The step of using the semantic recovery module to perform multi-scale fusion reconstruction of the semantic features, generating and outputting the original semantic signal, includes: The original semantic signal is generated using the following formula: ； in, S is the original semantic signal. β ( ) represents the semantic recovery module, D d ( ) represents the dynamic source-channel joint encoder. y represents the channel response, and y represents the distortion signal.

3. The channel model-driven semantic communication method in dynamic scenarios according to claim 1, characterized in that, The establishment of the network model, which includes a dynamic channel model, includes: A dynamic channel model with multiple inputs and multiple outputs is constructed. The dynamic channel model includes line-of-sight paths and non-line-of-sight paths characterized by multi-scatter cluster synthesis. The line-of-sight paths and the non-line-of-sight paths are used to reflect disturbances caused by mobility in the dynamic scenario.

4. The channel model-driven semantic communication method in dynamic scenarios according to claim 3, characterized in that, The disturbance includes Doppler frequency shift; The establishment of the network model includes: The mobile terminal and temporary base station in the dynamic scene are regarded as the relative receiving end and sending end; A uniform linear array antenna is configured at the receiving end and the transmitting end. The distance vector of the antenna and the Doppler frequency shift caused by the movement of the transmitting end are calculated.

5. The channel model-driven semantic communication method in dynamic scenarios according to claim 4, characterized in that, The calculation of the antenna's distance vector includes: The distance vector is calculated according to the following formula; ； in, Let Φ be the distance vector of the p-th antenna of the transmitting end. T Let θ be the angle of elevation. T For azimuth, δ T The distance between adjacent antennas.

6. The channel model-driven semantic communication method in dynamic scenarios according to claim 3, characterized in that, The disturbances include line-of-sight path loss, non-line-of-sight path loss, and small-scale fading. The line-of-sight path loss is calculated using the following formula: ； Among them, PL LoS For the line-of-sight path loss, d 3D Let d be the three-dimensional distance between the transmitter and the receiver. 2D f is the two-dimensional distance between the transmitter and the receiver. c The center frequency of the carrier. h is the distance between the breakpoints. BS h is the antenna height at the base station. UT The antenna height at the terminal; The non-line-of-sight path loss is calculated using the following formula: PL NLoS =max[PL Los ,13.54+39.08log 10 (d 3D )]+20log 10 (f c )-0.6(H R -1.5); Among them, PL NLoS H represents the non-line-of-sight path loss. R The antenna height of the receiving end; The channel matrix under line-of-sight path loss, non-line-of-sight path loss, and small-scale fading perturbations is calculated using the following formula: ； Where H(t, τ) is the channel matrix, PL[] is the path loss amplitude attenuation factor, and h pq (t, τ) represents the components of the small-scale fading. M R M represents the number of antennas at the receiving end. τ The number of antennas at the transmitting end.

7. The channel model-driven semantic communication method in dynamic scenarios according to claim 6, characterized in that, Before decoding the distorted signal using the channel impulse response provided by the trained dynamic channel model, the following steps are included: Calculate the distance vectors of multiple propagation paths between the receiver and the transmitter; The channel impulse response between an antenna at the receiving end and an antenna at the transmitting end is calculated based on multiple distance vectors.

8. The channel model-driven semantic communication method in dynamic scenarios according to claim 7, characterized in that, The calculation of the distance vectors of multiple propagation paths between the receiver and the transmitter includes: The propagation path is divided into the path from the transmitter to the first scatterer, the virtual link between the multiple scatterers, and the path from the last scatterer to the receiver. Calculate the time-varying distance vector from the transmitter to the first scatterer based on the moving speed of the transmitter; Calculate the stationary distance vector of the virtual link between the multiple scatterers, and the stationary distance vector from the last scatterer to the receiver.

9. The channel model-driven semantic communication method in dynamic scenarios according to claim 1, characterized in that, The trained network model includes: The network model is optimized using the following formula as the optimization objective: ； Among them, P dynamic This refers to the joint distribution of the source signal and channel response in the dynamic scenario described. D is a measure of semantic layer distortion. e For the dynamic source-channel joint encoder, D d For the dynamic source-channel joint decoder, z is the original dynamically refined signal, and h follows the dynamic joint distribution sampling. The output is the original semantic signal.

10. A channel model-driven semantic communication system for dynamic scenarios, characterized in that, The system includes: A module is established to build a network model; wherein, the network model includes a network architecture from the transmitter to the receiver and a dynamic channel model; A training module is used to train a network model and establish a communication system based on the trained network model. The communication system includes the trained network architecture. The network architecture includes a channel, a semantic extraction model located at the transmitting end, a semantic recovery model located at the receiving end, and a dynamic source-channel joint encoder and a dynamic source-channel joint encoder that combine a distance adaptive network from the transmitting end to the receiving end. The parameters of the channel are calculated from the trained dynamic channel model. An acquisition module is used to acquire the original dynamic scene signal at the sending end and extract semantic features from the original dynamic scene signal using the semantic extraction module. The encoding module is used at the transmitting end to take the channel response of the dynamic channel model as guiding information, and use the dynamic source-channel joint encoder to adaptively encode the semantic features to generate a semantic encoded signal. A transmission module is used to transmit the semantically encoded signal from the sending end to the receiving end. During the transmission process, the semantic encoding is distorted and becomes a distorted signal. The decoding module is used to receive the distorted signal using the receiving end, and decode the distorted signal using the channel impulse response provided by the trained dynamic channel model through the dynamic source-channel joint decoder to restore the semantic features; The output module is used to perform multi-scale fusion reconstruction of the semantic features using the semantic recovery module to generate and output the original semantic signal.

Images