A non-orthogonal multiple access word level power control method and system based on semantic importance

Abstract

The application discloses a non-orthogonal multiple access word level power control method and system based on semantic importance, and aims at the problems of mutual fragmentation of existing semantic transmission and non-orthogonal multiple access, coarse power allocation granularity and non-overlapping grouping, etc. User images are encoded into semantic word sequences by a semantic encoder, and normalized semantic importance is output by a self-attention module, so that the sum of all word importance is 1. According to the semantic importance, differentiated transmission power is allocated to each word, so that important words obtain higher power protection. The same user is allowed to overlap and join multiple NOMA groups, the remaining power is recovered, and diversity gain is obtained. With the joint target of minimizing semantic distortion and maximizing spectral rate, the grouping matrix, power share and codec parameters are cross-layer double-layer jointly optimized, the outer layer adopts multi-agent reinforcement learning to solve overlapping grouping and power allocation, and the inner layer adopts stochastic gradient descent to train the codec.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology in information science, specifically relating to a nonorthogonal multiple access term-level power control method based on semantic importance. Background Technology

[0002] In multi-user wireless image transmission systems, with the increasing number of terminal devices and the growing scarcity of channel resources, the traditional "source coding-channel coding-modulation" separate architecture struggles to guarantee image reconstruction quality under conditions of low signal-to-noise ratio and limited bandwidth. In recent years, the Deep Learning-based Joint Source-Channel Coding (DeepJSCC) method has demonstrated superior performance compared to traditional schemes in single-user point-to-point links by directly mapping images to channel symbols through an end-to-end neural network.

[0003] On the other hand, Non-Orthogonal Multiple Access (NOMA) technology allows multiple users to transmit over the same time-frequency resource block via power domain superposition, and then decode sequentially at the receiving end using Sequential Interference Cancellation (SIC), thereby improving spectral efficiency and the number of connections. Existing NOMA systems typically assume that each user belongs to only one packet, and that power is allocated as a whole on a user-by-user basis, with optimization targets mostly being traditional communication metrics such as bit rate or signal-to-noise ratio.

[0004] However, these two types of technologies have developed independently for a long time, lacking deep integration. Existing DeepJSCC methods do not consider power domain superposition, multi-user interference, and intra-group SIC decoding constraints in NOMA; while traditional NOMA methods treat the content to be transmitted as equally important bit sequences, lacking awareness of the importance of different semantic parts of the image. This results in bandwidth and power resources being evenly distributed between important and secondary semantics, limiting reconstruction quality. Furthermore, non-overlapping grouping limits the utilization of users' remaining transmit power and fails to capture the transmission opportunities and diversity gains provided by multiple independent fading resource blocks.

[0005] Therefore, there is an urgent need for a cross-layer joint optimization method that can combine semantic encoding, overlapping NOMA grouping, and refined power control to overcome the shortcomings of existing technologies. Summary of the Invention

[0006] This invention addresses the problems in existing technologies, such as the disconnect between semantic coding and multiple access, coarse power allocation granularity, non-overlapping grouping methods, and lack of cross-layer joint optimization modeling. It proposes a non-orthogonal multiple access term-level power control method based on semantic importance, aiming to minimize semantic distortion and maximize spectral rate, thereby improving the overall performance of multi-user image transmission systems.

[0007] The specific technical solution is as follows:

[0008] A nonorthogonal multiple access term-level power control method based on semantic importance includes: encoding the content to be transmitted into a semantic term sequence and extracting the semantic importance of each term, wherein the sum of the semantic importance of all terms is 1;

[0009] Based on the semantic importance, each lexical unit is assigned emission power, so that lexical units with high semantic importance receive higher emission power;

[0010] Users are assigned to one or more NOMA resource blocks to form NOMA groups. The same user is allowed to belong to multiple NOMA groups at the same time. Within each NOMA group, the token symbols of users belonging to that group are superimposed and transmitted in the power domain. The receiving end performs serial interference cancellation on the superimposed signal, decodes the token estimation, and reconstructs the content to be transmitted.

[0011] In some embodiments, the content to be transmitted is an image, which is encoded into a semantic word sequence by a semantic encoder, and the semantic importance of each word is output by the self-attention module at the end of the semantic encoder.

[0012] In step 2, the emission power of each word is determined according to the following formula:

[0013]

[0014] in, For user-level power scaling factor, For users In NOMA group The power share and satisfy , For semantic importance, This is the user's maximum transmit power.

[0015] In some embodiments, cross-layer joint optimization is also included with the joint objective of minimizing semantic distortion and maximizing spectral rate: the grouping matrix Δ, the user-level power scaling factor, etc. Group-level power share and semantic encoder With semantic decoder Using the parameters as variables, construct the following objective function:

[0016]

[0017] in And satisfy the serial interference elimination decodeability threshold constraint for each word; decompose the optimization problem into an outer combinatorial optimization subproblem and an inner encoding / decoding learning subproblem, and solve them iteratively.

[0018] In some embodiments, the outer combinatorial optimization subproblem uses a grouping matrix Δ and a user-level power scaling factor. Group-level power share For the decision variables, a multi-agent reinforcement learning algorithm based on a cooperative partially observable Markov decision process is used to solve them. Each user corresponds to one agent. After training under a centralized training-distributed execution framework, each agent makes independent decisions based on local observations during the online phase.

[0019] In some embodiments, the multi-agent reinforcement learning algorithm employs the QMIX algorithm: each agent's local Q-network takes local observations as input and outputs local action values; a centralized hybrid network generates non-negative weights based on the global state, monotonically aggregating the local action values ​​into a joint action value, satisfying... End-to-end training is performed using temporal difference loss.

[0020] When performing serial interference cancellation, the receiver determines the decoding order for each NOMA group according to the descending order of user channel gain within the group; for the current user to be decoded, its decoding residual is:

[0021]

[0022] in The received superimposed signal, This indicates that it takes precedence over the user in the serial interference cancellation sequence. Decoding users, The re-encoded symbols are obtained by inputting the reconstructed image of the decoded user back into its semantic encoder; the decoder directly uses the residual as input to learn implicit channel inversion through end-to-end training.

[0023] In some embodiments, a pass-through estimation is used to address the non-differentiable problem of serial interference elimination sorting: during forward propagation, the decoding order is determined according to channel estimation, and during backward propagation, the decoding order is regarded as a fixed arrangement, so that the gradient is back propagated to the semantic encoder parameters along the path of semantic encoder → semantic importance → transmit power → achievable rate; the semantic distortion loss includes mean square error loss and perception loss.

[0024] In some embodiments, the number of NOMA groups joined by each user satisfies The number of users within each NOMA group meets the following requirements. ,in Assign elements to the binary matrix Δ. The maximum number of groups a single user can join and ≥2, The maximum number of users within a single NOMA group; overlapping grouping allows the remaining transmit power of the same user to be recovered and utilized in multiple NOMA groups, and obtains diversity gain from multiple independent fading resource blocks.

[0025] In some embodiments, a lower limit on the emission power is set for lexical units with low semantic importance. This ensures that the signal-to-interference-plus-noise ratio at the receiving end is not lower than the preset serial interference cancellation decoding threshold. The lower limit of the transmission power Based on the minimum channel gain allowed by the system and the serial interference cancellation decoding threshold Pre-calculation;

[0026] The method is also applicable to downlink transmission scenarios, where the transmitter power budget constraint is the total transmit power of the base station. To replace per-user power budget in uplink scenarios The superposition coding order of each NOMA group is then used as an additional decision variable and solved together with the grouping matrix.

[0027] To implement the above method, the present invention also provides a communication system, including: a base station and multiple terminals communicating with the base station via a non-orthogonal multiple access (NOMA) protocol. The terminals are configured to: encode content to be transmitted into a semantic token sequence and extract the semantic importance of each token; allocate transmission power to each token according to its semantic importance; and allocate the tokens to one or more NOMA resource blocks to form NOMA groups. The same terminal may belong to multiple NOMA groups simultaneously. The token symbols of each terminal within each NOMA group are superimposed in the power domain and transmitted to the base station. The base station is configured to: receive the superimposed signal on each NOMA group; perform serial interference cancellation according to channel gain order; decode the token estimation; and reconstruct the content to be transmitted.

[0028] Preferably, the system further includes core network elements for storing and / or distributing semantic codec parameters or multi-agent reinforcement learning model parameters. The processors of the base station and the terminal are configured to implement the functions of the aforementioned steps.

[0029] Compared with the prior art, the present invention has at least the following technical effects:

[0030] 1. Significantly improved end-to-end image reconstruction quality. By assigning normalized semantic importance to each semantic unit generated from the user image encoding, and differentially allocating user transmit power at the unit granularity according to semantic importance, key semantic features in the image receive higher transmit power protection. Simultaneously, by incorporating the grouping matrix, user-level power scaling factor, group-level power share, and semantic codec parameters into a unified objective function for cross-layer, two-level joint optimization, semantic encoding / decoding becomes mutually compatible with multi-user NOMA access. These mechanisms work together to achieve consistent and significant improvements in PSNR and LPIPS compared to Orthogonal Multiple Access (OMA) schemes and non-overlapping NOMA schemes, especially in suppressing image blur artifacts caused by SIC residual interference in the low signal-to-noise ratio range.

[0031] 2. Improved Spectrum Efficiency and Resource Utilization Efficiency. Overlapping NOMA grouping allows the same user to join multiple NOMA groups simultaneously, enabling the recovery and utilization of the user's remaining transmit power and access to transmission opportunities and diversity gains from multiple independent fading resource blocks. Compared to non-overlapping NOMA schemes, both spectrum efficiency and system uplink achievable rate are improved. Furthermore, QMIX-based multi-agent reinforcement learning replaces exhaustive overlapping grouping search, reducing the solution complexity from exponential to linear. Online deployment requires only local Q-network forward inference, resulting in low latency and low signaling overhead. Moreover, the monotonicity constraint of QMIX ensures decision consistency between centralized training and distributed execution. Attached Figure Description

[0032] Figure 1: Schematic diagram of the semantic JSCC-NOMA system architecture of the present invention. (a) Traditional non-overlapping NOMA user grouping combined with DeepJSCC scheme; (b) Overlapping NOMA user grouping scheme proposed in this invention, in which the same user can use the remaining power to join multiple NOMA groups to obtain diversity gain.

[0033] Figure 2: Schematic diagram of the overlapping user grouping architecture based on QMIX in this invention, including distributed execution (local observation, agent Q network, independent argmax decision output grouping matrix) and centralized training (super network generates non-negative weights from global state, hybrid network aggregation). Monotonicity constraint It consists of two parts: TD loss end-to-end training. Detailed Implementation

[0034] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0035] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, an indirect connection through an intermediate medium, or the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0036] The overall idea of ​​this invention is to unify deep learning-based semantic coding, NOMA overlapping user grouping, word-level power allocation, and base station SIC decoding into a cross-layer joint optimization framework in a multi-user uplink wireless image transmission system. The overall idea specifically includes three points:

[0037] 1. On the user side, each image is mapped into a sequence of semantic tokens of fixed length by a semantic encoder. The self-attention module at the end of the encoder outputs a normalized semantic importance, so that each token has a quantifiable "weight" attribute, and thereby drives the differentiated allocation of power resources.

[0038] 2. On the multiple access side, a binary user-group allocation matrix is ​​introduced to achieve overlapping grouping, so that the same user can join multiple NOMA groups under the condition of satisfying grouping and decodability constraints, and grouping and group-level power share are used as decision variables.

[0039] 3. At the solution level, with the joint objective of minimizing semantic distortion and maximizing spectral rate, the overall optimization problem is modeled as a two-layer joint optimization problem with grouping and power share as outer variables and semantic codec as inner variables. The outer layer is solved by multi-agent reinforcement learning and the inner layer is solved by stochastic gradient descent. The two layers are iteratively coupled through reward backpropagation.

[0040] See Figures 1-2 The complete technical solution of the present invention consists of the following steps.

[0041] Step S1: System Modeling. The system consists of... The system consists of a single-antenna user and a base station (BS). The total system bandwidth... Divided into equal parts A number of mutually orthogonal NOMA resource blocks, each with a bandwidth of [missing information]. Establish a binary user-group allocation matrix. ,in Indicates user Join the The matrix must contain at least one NOMA group, otherwise it must contain 0. The matrix must satisfy constraint (C1): each user must join at least one group and at most one NOMA group. There are several groups (to achieve overlapping grouping), and a constraint (C2) that the number of users in each group does not exceed [number missing]. (Ensuring SIC feasibility). Each user uploads local images to the base station. .

[0042] Step S2: Semantic encoding and importance extraction (one of the core innovations). User Through the semantic JSCC encoder Encode the image into a length of [length]. semantic lexical sequence

[0043]

[0044] The semantic importance of each word is obtained by normalizing the attention logit from the self-attention module at the end of the encoder using softmax. And satisfy the following normalization constraints:

[0045]

[0046] Formula (1) is one of the core protection points of this invention. It refines the traditional bit-level and user-level power allocation unit to the semantic word granularity, so that different words within each user can obtain different transmit power according to their semantic importance. The normalization constraint ensures that the user-level power budget can be accurately divided into each word according to its importance without introducing additional power leakage.

[0047] Step S3: Lexicon-level power allocation, transmit signal generation, and uplink reception. Based on the semantic importance of the lexicon. Maximum transmit power of users User-level power scaling factor and users In its NOMA group Power share (satisfy ),user In the group The above is the first The emission power of each word is

[0048]

[0049] Forced lower bound on low importance lemmas This is to avoid violating the SIC decodable threshold. (User) In the group The first one sent on Each word metasignal is written as

[0050]

[0051] in This is a unit energy projection function. The base station is in the group. The received mixed signal is

[0052]

[0053] in For channel coefficients, The noise is additive white Gaussian noise. Within each group, each term must also satisfy the SIC decodeability constraint (denoted as C3), meaning that when performing SIC in descending order of channel gain, the received signal-to-interference-plus-noise ratio (SIR) of each term is not lower than a preset SIC decoding threshold. .

[0054] Step S4: Cross-layer dual-layer joint optimization modeling (second core innovation). This invention takes minimizing semantic distortion and maximizing spectral rate as the joint objective, and uniformly formulates the overall optimization problem of multi-user uplink semantic transmission as the following joint optimization problem:

[0055] Objective function:

[0056]

[0057] Constraints:

[0058]

[0059] in These are weighting coefficients. This is a power soft penalty term during the training phase. For users semantic distortion loss, Given its uplink reachable rate. Constraints (C1) and (C2) ensure the validity of the grouping structure (upper and lower bounds of overlapping groups and upper bound of group size); (C3) ensures that the SIC of each word is decodable.

[0060] Formula (2) is the second core protection point of this invention. It will group the matrix. User-level power scaling factor Group-level power share and semantic encoder With decoder Joint modeling under the same objective function unifies the multi-user access of the physical layer and the image reconstruction of the semantic layer into a cross-layer joint optimization problem, which is the mathematical basis for the subsequent coupling and iterative solution of outer-layer combinatorial optimization and inner-layer learning.

[0061] Based on this modeling, the original problem is decomposed into two mutually coupled but non-overlapping subproblems:

[0062] Outer combinatorial optimization subproblem: , , Given the decision variables, solve for the optimal grouping and power share;

[0063] Inner layer encoding / decoding learning sub-problem: with semantic encoders With decoder Using these as decision variables, we solve for the optimal semantic codec network parameters.

[0064] After each policy update by the outer layer, the current policy will be updated. The result is passed to the inner layer; after the inner layer completes a gradient update based on this, it passes the obtained reward back to the outer layer, thus realizing the two layers iteratively solving the problem.

[0065] Step S5: Solving the outer layer's overlapping grouping and power share based on MARL. The outer layer combinatorial optimization subproblem is modeled as a cooperative partially observable Markov decision process, with one agent corresponding to each user. The QMIX value function decomposition algorithm is used, and training is performed under a centralized training-distributed execution (CTDE) framework: each agent independently estimates the local action value based on local observations. Centralized hybrid networks, conditioned on the global state, generate non-negative weights through a hypernetwork, and integrate the various... Monotonic aggregation as the value of joint action And satisfy the monotonicity constraint. The overall network is trained end-to-end using temporal difference loss. During the online phase, only the local Q-networks of each agent are deployed. Based on their own observations, each agent selects candidate NOMA action groups. All action combinations are then subjected to feasibility screening by the environment against pairs (C1), (C2), and (C3) to obtain the final grouping matrix. With the corresponding , This step is a tool-like step in solving the outer subproblem of formula (2) using MARL, and the specific implementation adopts the existing QMIX algorithm.

[0066] Step S6: End-to-end SGD training of the inner semantic codec. Given the outer layer... Below, for each user Minimize semantic distortion loss using SGD

[0067]

[0068] Group SIC within the group by user channel gain Proceed sequentially. For the current user to be decoded... Its residual is

[0069]

[0070] in Indicates that it precedes in the SIC order. Decoding users, For users who have already decoded The recoded symbols, from which the image is reconstructed Re-enter into encoder The decoder directly uses the residual as input: It relies on end-to-end training to learn implicit channel inversion. For the non-differentiable problem of SIC ordering, a "straight-through estimation" method is used: the forward pass determines the decoding order based on the current channel estimate, and the backward pass is treated as a fixed arrangement, causing the gradient to move along... Return the encoder parameters.

[0071] Step S7: Base Station SIC Decoding and Image Reconstruction. The base station performs SIC sequentially on each group of received mixed signals according to the above residual recursion to obtain semantic token estimates for each user. Then through the semantic JSCC decoder Complete image reconstruction The semantic distortion loss is constructed by MSE and LPIPS, and the parameters related to encoder-decoder and power control are updated back, thus completing the entire cross-layer joint training closed loop.

[0072] Step S8: Extended Explanation of Downlink Scenarios. Steps S1–S7 above are described using multi-user uplink image transmission as a typical scenario. The cross-layer joint optimization framework combining semantic coding, overlapping NOMA grouping, and term-level power control proposed in this invention can also be extended to downlink scenarios, with the following two structural differences from the uplink scenario. First, the power budget constraint at the transmitting end changes: In the downlink scenario, the base station uniformly performs semantic coding on the images of all users, and in each NOMA group… After superimposing and encoding the token streams of the group of users, the data is broadcast to each user on the corresponding resource block; at this point, the maximum transmit power of a single user is no longer used. As a constraint, the total transmit power of the base station As a constraint, namely First, it replaces the per-user power budget in the original uplink formula. Second, the intra-group superposition coding order becomes a new optimization variable: In the uplink scenario, the base station performs SIC in descending order of the instantaneous channel gain of users within the group, and the decoding order is directly determined by the current channel estimation; while in the downlink scenario, each NOMA group... superposition encoding order It itself needs to be used as a decision variable and grouping matrix User-level power scaling factor and group-level power share Solve them together, among which Indicates in group In the superposition encoding, it is in the first The user index is 1 bit. The remaining steps (lexical extraction, lexical-level power allocation mapping, cross-layer bilayer joint optimization modeling, and MARL-based solution framework) remain consistent in the downlink scenario, only requiring the action space of the outer MARL to be expanded accordingly to cover the selection of the superposition encoding order.

[0073] In specific network element deployment examples, the offline training phase is conducted on the core network side or a dedicated server. The semantic encoder, decoder, and QMIX network parameters obtained from the training are distributed to each terminal and base station via the radio access network through core network elements (such as UPF or SMF). During online operation, the semantic encoder module and power allocation module of the terminal perform word-level power allocation, and the resource allocation and scheduling module of the base station runs multi-agent reinforcement learning inference to determine the grouping matrix and power share. Finally, the serial interference cancellation module and semantic decoder module of the base station complete the image reconstruction, thereby achieving end-to-end data flow closed loop.

[0074] The following detailed explanation will be provided using a specific example:

[0075] (a) Implementation Environment

[0076] Hardware: Servers or workstations containing one or more GPUs for offline centralized training; lightweight local Q networks and semantic encoders can be deployed on the user side.

[0077] Software: A deep learning framework based on PyTorch, including the NAFNet backbone network as the semantic JSCC encoder and decoder, a self-attention module for outputting lexical importance, and the QMIX multi-agent reinforcement learning module.

[0078] Channel model: Uplink Rayleigh flat fading, additive white Gaussian noise.

[0079] Dataset: CelebA Face Image Dataset.

[0080] (II) Implementation Steps

[0081] 1. To Individual users Initialize the uplink scenario for each NOMA resource block; set the maximum number of groups each user can join. Maximum number of users per group SIC threshold Lower limit of word power .

[0082] 2. User Image Input Semantic JSCC Encoder , thus obtaining the word sequence Importance And perform normalization verification according to formula (1).

[0083] 3. The outer QMIX network outputs candidate group actions for each agent based on the current local observations. After the environment performs feasibility screening (C1)(C2)(C3), the action for this time slot is obtained. Simultaneously output by MARL and .

[0084] 4. User press Calculate the transmit power of each word and send it. .

[0085] 5. The base station performs SIC in descending order of channel gain on each group, and recursively calculates it based on the residual. Decoding and reconstruct the image .

[0086] 6. Calculate the MSE and LPIPS losses, construct the sample loss according to the objective function of formula (2), and perform SGD update on the inner codec; return the obtained reward to the outer QMIX for temporal difference update.

[0087] 7. Repeat steps 2–6 until the target values ​​of formula (2) (outer layer reward and inner layer distortion loss) converge.

[0088] 8. For the implementation of the downlink scenario, replace the transmitter power budget constraint as described in step S8 (based on the total transmit power of the base station). Replace per-user budget And the superposition coding order of each NOMA group The action space of the outer MARL is optimized together, and the remaining encoding, normalization, word-level power mapping, SIC / per-user decoding and codec update process are consistent with the above uplink steps.

[0089] (III) Key Parameters

[0090] Wireless communication environment parameters:

[0091] Cell radius m;

[0092] Bandwidth per NOMA resource block kHz;

[0093] Uplink user maximum transmit power budget dBm;

[0094] White noise power spectral density dBm / Hz;

[0095] Both the base station and the user terminal have a single antenna.

[0096] MARL Network and Training Parameters:

[0097] The local Q network has a hidden layer dimension of 128; the super network has a hidden layer dimension of 64.

[0098] Using the Adam optimizer, the learning rate... ;

[0099] Perceptual weights in inner semantic distortion loss It is a positive value.

[0100] (iv) Implementation Results

[0101] In the CelebA dataset and Rayleigh flat fading channel (channel gain difference) dB, number of users Below, this scheme is compared with orthogonal multiple access (OMA) baselines and non-overlapping NOMA baselines in terms of PSNR and LPIPS:

[0102] In the low to medium reference signal-to-noise ratio range, this scheme significantly outperforms the OMA and non-overlapping NOMA baselines in both PSNR and LPIPS.

[0103] In number of users Within the specified range, the image reconstruction performance of this solution remains stable and consistently outperforms other solutions.

[0104] Compared with non-overlapping NOMA, the overlapping grouping in this scheme can effectively suppress image blur artifacts caused by SIC residual interference, and the LPIPS index is improved more significantly.

[0105] The above results correspond one-to-one with the two core protection points of this invention:

[0106] Power allocation driven by word-level semantic importance normalization in formula (1): By finely focusing power on high-importance words, key semantic features achieve higher equivalent signal-to-noise ratios, thereby bringing about simultaneous improvements in LPIPS and PSNR (especially noticeable compared to the non-overlapping NOMA baseline with uniform power allocation per user).

[0107] The cross-layer dual-layer joint optimization modeling attributed to Equation (2) is achieved by alternately solving the grouping matrix, power share and semantic codec in the same objective function, which makes the semantic codec and multi-user access mutually compatible, and obtains a significant image reconstruction quality gain relative to the OMA baseline designed independently by each module.

[0108] The above embodiments illustrate in detail the specific implementation of the technical solution of the present invention, the logical and connection relationships of each component, and the complete working process. Those skilled in the art will understand that various changes and modifications can be made to the above embodiments without departing from the principles and spirit of the present invention, and all such changes and modifications should fall within the protection scope of the appended claims.

Claims

1. A non-orthogonal multi-access term-level power control method based on semantic importance, characterized in that, include: The content to be transmitted is encoded into a semantic word sequence and the semantic importance of each word is extracted. The sum of the semantic importance of all words is 1. Based on the semantic importance, each lexical unit is assigned emission power, so that lexical units with high semantic importance receive higher emission power; Users are assigned to one or more NOMA resource blocks to form NOMA groups. The same user is allowed to belong to multiple NOMA groups at the same time. Within each NOMA group, the token symbols of users belonging to that group are superimposed and transmitted in the power domain. The receiving end performs serial interference cancellation on the superimposed signal, decodes the token estimation, and reconstructs the content to be transmitted.

2. The method according to claim 1, characterized in that, The content to be transmitted is an image. The image is encoded into a semantic word sequence by a semantic encoder, and the semantic importance of each word is output by the self-attention module at the end of the semantic encoder. The emission power of each word is determined according to the following formula: in, For user-level power scaling factor, For users In NOMA group The power share and satisfy , For semantic importance, This is the user's maximum transmit power.

3. The method according to claim 1, characterized in that, It also includes cross-layer joint optimization with the joint objective of minimizing semantic distortion and maximizing spectral rate: combining the grouping matrix Δ and the user-level power scaling factor. Group-level power share and semantic encoder With semantic decoder Using the parameters as variables, construct the following objective function: in And satisfy the serial interference elimination decodeability threshold constraint for each word; decompose the optimization problem into an outer combinatorial optimization subproblem and an inner encoding / decoding learning subproblem, and solve them iteratively.

4. The method according to claim 3, characterized in that, The outer-layer combinatorial optimization subproblem uses the grouping matrix Δ and the user-level power scaling factor. Group-level power share For the decision variables, a multi-agent reinforcement learning algorithm based on a cooperative partially observable Markov decision process is used to solve them. Each user corresponds to one agent. After training under a centralized training-distributed execution framework, each agent makes independent decisions based on local observations during the online phase.

5. The method according to claim 4, characterized in that, The multi-agent reinforcement learning algorithm employs the QMIX algorithm: each agent's local Q-network takes local observations as input and outputs local action values; the centralized hybrid network generates non-negative weights based on the global state, monotonically aggregating the local action values ​​into a joint action value, satisfying... End-to-end training is performed using temporal difference loss.

6. The method according to claim 1, characterized in that, When performing serial interference cancellation, the receiver determines the decoding order for each NOMA group according to the descending order of user channel gain within the group; for the current user to be decoded, its decoding residual is: in The received superimposed signal, This indicates that it takes precedence over the user in the serial interference cancellation sequence. Decoding users, The re-encoded symbols are obtained by inputting the reconstructed image of the decoded user back into its semantic encoder; the decoder directly uses the residual as input to learn implicit channel inversion through end-to-end training.

7. The method according to claim 6, characterized in that, The non-differentiable problem of serial interference elimination sorting is handled by direct-pass estimation: during forward propagation, the decoding order is determined according to channel estimation, and during backward propagation, the decoding order is regarded as a fixed arrangement, so that the gradient is back propagated to the semantic encoder parameters along the path of semantic encoder → semantic importance → transmit power → achievable rate; the semantic distortion loss includes mean square error loss and perception loss.

8. The method according to claim 1, characterized in that, The number of NOMA groups joined by each user satisfies The number of users within each NOMA group meets the following requirements. ,in Assign elements to the binary matrix Δ. The maximum number of groups a single user can join and ≥2, The maximum number of users within a single NOMA group; overlapping grouping allows the remaining transmit power of the same user to be recovered and utilized in multiple NOMA groups, and obtains diversity gain from multiple independent fading resource blocks.

9. The method according to claim 1, characterized in that, Set a lower limit on emission power for lemmas with low semantic importance. This ensures that the signal-to-interference-plus-noise ratio at the receiving end is not lower than the preset serial interference cancellation decoding threshold. The lower limit of the transmission power Based on the minimum channel gain allowed by the system and the serial interference cancellation decoding threshold Pre-calculation; The method is also applicable to downlink transmission scenarios, where the transmitter power budget constraint is the total transmit power of the base station. To replace per-user power budget in uplink scenarios The superposition coding order of each NOMA group is then used as an additional decision variable and solved together with the grouping matrix.

10. A communication system for multi-user wireless transmission, characterized in that, include: A base station, and multiple terminals communicating with the base station via a non-orthogonal multiple access (NOMA) protocol; The terminal is used to: encode the content to be transmitted into a semantic word sequence and extract the semantic importance of each word, the sum of the semantic importance of all words being 1; allocate transmission power to each word according to the semantic importance, and allocate it to one or more NOMA resource blocks to form a NOMA group. The same terminal belongs to multiple NOMA groups at the same time, and the word symbols of each terminal in each NOMA group are superimposed in the power domain and sent to the base station. The base station is used to: receive superimposed signals on each NOMA group, perform serial interference cancellation in the order of channel gain, decode the token estimation, and reconstruct the content to be transmitted.

Images