Multi-user satellite-ground communication method based on position semantic information and communication system thereof
By adopting a multi-user satellite-ground communication method based on location semantic information in a low signal-to-noise ratio environment, semantic encoding is performed using multi-head attention mechanism and relative position information, and channel encoding is performed in combination with prior channel knowledge, the problem of insufficient accuracy of semantic information exchange transmission is solved, and higher semantic feature representation accuracy and system robustness are achieved.
Patent Information
- Application Number
- CN202411973754.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-15
AI Technical Summary
In a low signal-to-noise environment, the transmission accuracy of semantic-level information exchange has not yet reached an ideal level, especially in low-earth orbit satellite communications, problems such as dynamic topology, high connection density and severe signal fading have weakened signal strength and reduced service quality.
The multi-user satellite-ground communication method based on location semantic information is adopted, and the semantic coding of the satellite end and the user end decoding method is used to semantic coding using the multi-head attention mechanism and relative position information, and channel coding is combined with prior channel knowledge to achieve efficient extraction of semantic features and accurate signal transmission.
By introducing word-level relative position information, encoding and decoding can accurately capture semantic dependencies, improve the accuracy of semantic feature representation, and enhance the robustness and reliability of the system. The simulation results show that BLEU and sentence similarity performance indicators are better than the existing technology.
Smart Images
Figure CN119834852A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a multi-user satellite-to-ground communication method and a communication system thereof based on position semantic information, and belongs to the technical field of information and communication. Background Art
[0002] The next generation of communication networks is developing towards higher data rates, wider coverage, and diversified application scenarios. However, due to the complexity of maintenance and high costs, the existing ground communication system only covers about 20% of the land area, which cannot meet the growing global communication needs. Low Earth Orbit (LEO) satellite communication has become one of the key technologies to break through this bottleneck with its wide-area coverage and cost-effectiveness. Despite its significant advantages, LEO satellite communication still needs to overcome technical difficulties such as dynamic topology, high connection density, and severe signal fading, which will significantly weaken the signal strength and reduce the quality of service. In addition, limited resources and low signal-to-noise ratio (SNR) environment further exacerbate these challenges, making it difficult for traditional technologies to be efficiently applied in this context. Addressing these complex issues requires pioneering technologies and innovations in communication paradigms.
[0003] Weaver proposed that communication can be divided into three levels: the syntactic level emphasizes the accuracy of symbol transmission, the semantic level focuses on the transmission of information meaning, and the pragmatic level focuses on the actual impact of information on behavior. Traditional communication technology usually stays at the syntactic level, aiming at accurate symbol transmission, but its transmission efficiency is difficult to meet the needs of modern communication. In order to break through this limitation, the research focus has gradually shifted to the semantic level, changing the core of transmission from symbols to semantic information. This development has promoted the rise of semantic communication and become a new paradigm of communication technology. Semantic communication achieves higher compression rate, noise resistance and receiving end effect by extracting and transmitting key information. This method has attracted widespread attention in the academic community, especially in the application of low-Earth orbit satellite communications, showing its great potential in complex scenarios. However, in low signal-to-noise ratio environments, the transmission accuracy of semantic-level information exchange needs to be improved. Summary of the invention
[0004] In order to improve the accuracy of semantic-level information exchange and transmission, the present invention provides a multi-user satellite-to-ground communication method and a communication system based on location semantic information.
[0005] A multi-user satellite-to-ground communication method based on position semantic information of the present invention comprises a satellite-side encoding method and a user-side decoding method;
[0006] Satellite encoding methods include:
[0007] Combine sentences into a matrix S;
[0008] The matrix S is semantically encoded based on the multi-head attention mechanism and converted into a semantic feature matrix M. The multi-head attention mechanism constructs an attention score based on the relative position information between words.
[0009] The semantic feature matrix M is combined with the prior channel knowledge vector to perform channel coding to generate the signal matrix X. After the signal matrix X is transmitted via the satellite-ground channel, the signal Y received by user k is k , k = 1, 2…K, K represents the number of users;
[0010] The user-side decoding methods include:
[0011] The received signal Y k Restore to semantic feature matrix
[0012] The semantic feature matrix Convert to sentence matrix
[0013] Sentence Matrix Separate and obtain the sentence content of the corresponding user.
[0014] As a preference, the attention score for:
[0015]
[0016] Among them, S q and S k is the query matrix and key matrix obtained from the matrix S, W q YesS q The projection matrix, W k,E and W k,P They are S k Corresponding to independent content-based and position-based projection matrices, E i and E j are the hidden layer matrices of input word i and word j respectively, v is the dynamically updated trainable weight, P i-j Represents the relative distance between word i and word j;
[0017] l max Indicates the maximum relative distance.
[0018] As a preference, the matrix S is input into the multi-head attention layer to obtain the representative semantic features S in the matrix S m :
[0019] S m =Concat(head1,...,head l )Wo
[0020] Among them, Concat(·) represents the vector concatenation operation, W o is the output layer weight matrix of the multi-head attention layer;
[0021] S v is the value matrix obtained from the matrix S, d represents the hidden layer dimension of the multi-head attention layer, and Softmax(·) represents the softmax function;
[0022] For the semantic feature S m Perform residual connection and normalize the result after residual connection. After normalization, output feature S r :
[0023] S r =Normalization(S e +S m )
[0024] Among them, Normalization(·) represents the layer normalization operation, S e represents the embedding vector;
[0025] The feature S r Map to high-dimensional space, and then map back to low-dimensional space to generate feature S d ;
[0026] For feature S d Perform residual connection and normalize the result after residual connection. After normalization, output the semantic feature matrix M:
[0027] M=Normalization(S r +S d ).
[0028] As a preference, the feature S r Methods for mapping to high-dimensional space and then back to low-dimensional space include:
[0029] The feature S r Input to the first linear layer, the output of the first linear layer
[0030]
[0031] ReLU(·) is the activation function, W1 and b1 are the weight matrix and bias vector of the first linear layer respectively;
[0032] Will Input to the second linear layer, the output feature S of the second linear layer d:
[0033]
[0034] W2 and b2 are the weight matrix and bias vector of the second linear layer respectively.
[0035] Preferably, the user-side decoding method further comprises:
[0036] The sentence matrix Input to the softmax layer, the softmax layer is used to generate a probability distribution for all possible words, and its calculation formula is:
[0037]
[0038] Among them, z i,j is the sentence matrix The i-th element in the j-th row, I c is the number of categories.
[0039] Preferably, the method of combining the semantic feature matrix M with the priori channel knowledge vector to perform channel coding to generate the signal matrix X includes:
[0040] The semantic feature matrix M and the prior channel knowledge vector D are embedded and concatenated to generate discrete features.
[0041]
[0042] Where, Embedding(·) represents the embedding and concatenation operation, and D is the vector formed by concatenating the LEO satellite velocity and signal-to-noise ratio;
[0043] According to discrete features Get the signal matrix X:
[0044]
[0045] Where Dense(·) represents dense block operation, Sigmoid(·) represents sigmoid function, W3 and b3 are trainable weights and biases, respectively.
[0046] The present invention also provides a communication system of the multi-user satellite-to-ground communication method based on location semantic information, comprising a location-aware semantic encoder, a knowledge-assisted channel encoder, a knowledge-assisted channel decoder, a location-aware semantic decoder and a multi-user discriminator;
[0047] A location-aware semantic encoder and a knowledge-assisted channel encoder are arranged at the satellite end, and a knowledge-assisted channel decoder, a location-aware semantic decoder and a multi-user discriminator are arranged at the user end;
[0048] A position-aware encoder is used to semantically encode the matrix S based on a multi-head attention mechanism and convert it into a semantic feature matrix M. The multi-head attention mechanism constructs an attention score based on the relative position information between words.
[0049] The knowledge-assisted encoder is used to combine the semantic feature matrix M with the prior channel knowledge vector to perform channel coding and generate the signal matrix X;
[0050] After the signal matrix X is transmitted via the satellite-ground channel, the signal Y received by user k is k , k = 1, 2…K, K represents the number of users;
[0051] A knowledge-assisted decoder is used to convert the received signal Y k Restore to semantic feature matrix
[0052] The position-aware decoder is used to transform the semantic feature matrix into Convert to sentence matrix
[0053] Multi-user discriminator for sentence matrix Separate and obtain the sentence content of the corresponding user.
[0054] The beneficial effects of the present invention are as follows: the present invention transmits multiple user signals simultaneously through the satellite end, and efficiently distinguishes them based on semantic features. A position-aware semantic encoding and decoding method is designed to extract user semantic information and reduce interference between users. By introducing word-level relative position information, the encoding and decoding can accurately capture semantic dependencies and improve the accuracy of semantic feature representation. At the same time, the knowledge-assisted channel encoding and decoding method makes full use of prior channel knowledge to enhance the robustness and reliability of the system. Simulation results show that the present invention is superior to the prior art in bilingual evaluation score (BLEU) and sentence similarity performance indicators, which fully demonstrates its excellent performance in low-orbit satellite communication applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a network structure diagram of the multi-user satellite semantic communication method proposed by the present invention;
[0056] Figure 2 It is the algorithm convergence curve diagram of the training and verification process;
[0057] Figure 3 is a BLEU performance comparison chart of different benchmark solutions and the present invention, wherein SatSC represents the solution of the present invention;
[0058] Figure 4It is a comparison chart of sentence similarity performance between different benchmark schemes and the present invention. DETAILED DESCRIPTION
[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0060] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.
[0062] The multi-user satellite-to-ground communication method based on location semantic information of this embodiment is a multi-user satellite semantic communication method. It takes advantage of information compression in semantic communication and adopts a multi-user and satellite semantic communication scheme to accurately transmit text information of multiple users. And a new semantic coding and channel coding are designed to improve the accuracy of text transmission between satellite and ground systems. Semantic coding takes advantage of relative position coding to process multi-user semantic features and accurately extract semantic information of multiple users. Channel coding uses channel prior knowledge to adapt to complex satellite channel environments.
[0063] The multi-user satellite-to-ground communication method based on location semantic information of this embodiment includes an encoding method implemented at the satellite end and a multi-user decoding method, and can also be implemented in the form of a communication system, combined with Figure 1 This embodiment is introduced in the form of a communication system. The communication system of this embodiment can be referred to as the SatSC architecture, which includes a location-aware semantic encoder, a knowledge-assisted channel encoder, a knowledge-assisted channel decoder, a location-aware semantic decoder, and a multi-user discriminator to improve the performance and robustness of multi-user text transmission. Specifically, the location-aware semantic codec can capture the deep semantic relationship between words and efficiently extract the independent semantic features of each user by embedding word-level relative position information into the semantic encoding process. In addition, the knowledge-assisted channel codec uses prior channel knowledge to effectively reduce the interference effect of the satellite channel on signal transmission during the encoding and decoding process, thereby enhancing the system's anti-interference capability.
[0064] In the data transmission stage, each sentence is first padded to a uniform length L and combined into a matrix Where B represents the batch size and K is the number of users. Subsequently, the location-aware semantic encoder transforms the matrix S into a semantic feature matrix based on a multi-head attention mechanism. Where V is the vector dimension of the semantic feature. Then, the knowledge-assisted channel encoder combines the semantic feature matrix M with the prior channel information to generate the signal matrix Where N represents the output dimension of the channel encoder. After the signal X is transmitted via the satellite-terrestrial channel, the signal received by each user is
[0065] In the decoding stage, the knowledge-assisted channel decoder uses a structure symmetrical to the channel encoder to convert the received signal Y k Restore to semantic feature matrix Then, the position-aware semantic decoder uses a structure symmetrical to the channel encoder to convert Convert to sentence matrix It contains the sentences of all users. Finally, the ALBERT model can be used as a multi-user discriminator to accurately separate the sentence content of each user, thereby completing the entire process of multi-user text transmission. This implementation effectively combines semantic feature modeling with channel characteristics, significantly improving the performance of the system in complex satellite channel environments.
[0066] In order to incorporate the dependencies between words into the semantic encoder-decoder, this implementation proposes a relative position attention mechanism to replace the original attention score in the multi-head attention mechanism. Looking back at the original attention mechanism, its expression is as follows:
[0067]
[0068] Where d represents the dimension of the hidden layer, S q , S k and S v They are the query matrix, key matrix, and value matrix respectively.
[0069] Obviously, the attention score However, this method only contains the initial position information of the word and fails to make full use of the relative position information between words. i,j Replaced with relative position attention score
[0070] This implementation Integrates word-level relative position information P i-j And trainable weights v to enhance the dependencies between words, the specific form is:
[0071]
[0072] Among them, W q YesS q The projection matrix, W k,E and W k,P YesS k Corresponding to independent content-based and position-based projection matrices, E i and E j is the input hidden layer matrix, v is the dynamically updated trainable weight, P i-j represents the relative distance between word i and word j, which is defined as:
[0073]
[0074] Among them, l max Represents the maximum relative distance. By introducing the relative position attention mechanism, this implementation can more effectively capture the dependency relationship of basic units in the transmission sequence, thereby significantly improving the semantic encoder's ability to model the semantic association between words. This improvement provides greater robustness and flexibility for semantic communication in complex channel environments.
[0075] Based on this, the attention mechanism in the multi-head attention layer introduces the relative position attention score It achieves the goal of giving greater weight to vectors with semantic dependencies, thereby generating more effective semantic representations. Each attention head focuses on a specific semantic relationship and generates semantic representations of different dimensions. By integrating the outputs of multiple attention heads, a representative semantic feature S can be obtained. m , whose expression is:
[0076]
[0077] Among them, Concat(·) represents the vector concatenation operation, l is the number of attention heads, S m is the output of the multi-head attention layer, W o is the weight matrix of the output layer in the multi-head attention layer.
[0078] Between the multi-head attention layer and the dense block, a residual connection is introduced to preserve feature information and improve gradient fluidity. Subsequently, the output after the residual connection is layer normalized, and its expression is:
[0079] S r =Normalization(S e +S m ) (5)
[0080] Among them, Normalization(·) represents the layer normalization operation. The normalized output S r It will be passed to the dense block to further extract more discriminative feature representations.
[0081] The dense block extracts deep features through two linear transformations and subsequent nonlinear activation functions. Specifically, the dense block first transforms S r Mapped to a high-dimensional space, and then mapped back to a low-dimensional space, finally generating a feature representation S d , the calculation process is as follows:
[0082]
[0083] in, is the output of the first linear layer, ReLU(·) is the activation function, W1 and b1 are the weight matrix and bias vector of the first linear layer, W2 and b2 are the weight matrix and bias vector of the second linear layer. Finally, the feature matrix S d After residual connection and normalization layer, the semantic representation M is generated, which is the output of the position-aware semantic encoder. The complete process is:
[0084] M=Normalization(S r +S d ) (8)
[0085] The position-aware semantic decoder is the inverse operation of the semantic encoder. Specifically, the input semantic vector It first passes through a multi-head attention layer with a word-level relative position encoding scheme, and then a normalization layer integrated with residual connections. In the subsequent feature extraction stage, the intermediate feature representation is further optimized through dense blocks to reconstruct the semantic information.
[0086] In addition, the output of the semantic decoder of this embodiment also integrates a softmax layer to convert the sentence matrix Input to the softmax layer, the softmax layer is used to generate a probability distribution for all possible words, and its calculation formula is:
[0087]
[0088] Among them, z i,j is the sentence matrix The i-th element in the j-th row, I c is the number of categories.
[0089] In order to make full use of the prior satellite channel knowledge, the knowledge-assisted channel encoder of this embodiment first embeds and concatenates the semantic feature matrix M output by the semantic encoder and the knowledge vector D of the prior channel to generate a discrete feature representation. The process is as follows:
[0090]
[0091] Where, Embedding(·) represents the embedding and concatenation operation, D is the vector formed by concatenating the LEO satellite velocity and signal-to-noise ratio (SNR), is the feature representation of the output of the embedding layer.
[0092] Next, is passed to the conditional block to extract channel features. As the core module that integrates prior satellite channel knowledge into the channel codec, the conditional block consists of dense blocks and neural network layers to extract potential features from prior satellite channel knowledge. Specifically, the conditional block generates channel features through discrete feature representation and adaptively adjusts the importance of prior knowledge. Its expression is:
[0093]
[0094] Where Dense(·) represents a dense block operation, Sigmoid(·) represents the sigmoid function, X is the output of the channel encoder, and W3 and b3 are trainable weights and biases, respectively.
[0095] This implementation can take advantage of the semantic encoder to solve the multi-user feature extraction problem, and use the channel encoder to enhance the adaptability to the satellite channel. By using relative position information and channel information, the ability to effectively transmit text under the satellite channel is obtained, thereby improving the transmission accuracy; simulation results show that compared with traditional and deep learning-based solutions, this implementation significantly improves transmission accuracy.
[0096] The present invention considers combining semantic communication with satellite-to-ground communication systems, providing a new solution for improving communication performance. In a low signal-to-noise ratio environment, semantic-level information exchange can effectively improve transmission accuracy. In addition, the global coverage advantage of LEO satellites makes the updating and synchronization of semantic background knowledge more efficient, thereby enhancing the reliability and scalability of the system. This fusion not only provides technical support for global seamless connectivity, but also significantly optimizes transmission efficiency and robustness, and experimentally verifies the expected performance enhancement of the satellite network semantic communication model.
[0097] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.
Claims
1. A multi-user satellite-to-ground communication method based on location semantic information, characterized in that: It includes a satellite end encoding method and a user end decoding method; The satellite encoding method comprises: Combine sentences into a matrix S; The matrix S is semantically encoded based on the multi-head attention mechanism and converted into a semantic feature matrix M. The multi-head attention mechanism constructs an attention score based on the relative position information between words. The semantic feature matrix M is combined with the prior channel knowledge vector to perform channel coding to generate the signal matrix X. After the signal matrix X is transmitted via the satellite-ground channel, the signal Y received by user k is k , k = 1, 2…K, K represents the number of users; The user terminal decoding method comprises: The received signal Y k Restore to semantic feature matrix The semantic feature matrix Convert to sentence matrix Sentence Matrix Separate and obtain the sentence content of the corresponding user.
2. The multi-user satellite-to-ground communication method based on location semantic information according to claim 1, characterized in that: Attention score for: Among them, S q and S k is the query matrix and key matrix obtained from the matrix S, W q YesS q The projection matrix, W k,E and W k,P They are S k Corresponding to independent content-based and position-based projection matrices, E i and E j are the hidden layer matrices of input word i and word j respectively, v is the dynamically updated trainable weight, P i-j Represents the relative distance between word i and word j; l max Indicates the maximum relative distance.
3. The multi-user satellite-to-ground communication method based on location semantic information according to claim 1, characterized in that: Input the matrix S into the multi-head attention layer to obtain the representative semantic features S in the matrix S m : S m =Concat(head1,...,head l )W o Among them, Concat(·) represents the vector concatenation operation, W o is the output layer weight matrix of the multi-head attention layer; S v is the value matrix obtained from the matrix S, d represents the hidden layer dimension of the multi-head attention layer, and Softmax(·) represents the softmax function; For the semantic feature S m Perform residual connection and normalize the result after residual connection. After normalization, output feature S r : S r =Normalization(S e +S m ) Among them, Normalization(·) represents the layer normalization operation, S e represents the embedding vector; The feature S r Map to high-dimensional space, and then map back to low-dimensional space to generate feature S d ; For feature S d Perform residual connection and normalize the result after residual connection. After normalization, output the semantic feature matrix M: M=Normalization(S r +S d )。 4. The multi-user satellite-to-ground communication method based on location semantic information according to claim 1, characterized in that: The feature S r Methods for mapping to high-dimensional space and then back to low-dimensional space include: The feature S r Input to the first linear layer, the output of the first linear layer ReLU(·) is the activation function, W1 and b1 are the weight matrix and bias vector of the first linear layer respectively; Will Input to the second linear layer, the output feature S of the second linear layer d : W2 and b2 are the weight matrix and bias vector of the second linear layer respectively.
5. The multi-user satellite-to-ground communication method based on location semantic information according to claim 1, characterized in that: The user terminal decoding method also includes: The sentence matrix Input to the softmax layer, the softmax layer is used to generate a probability distribution for all possible words, and its calculation formula is: Among them, z i,j is the sentence matrix The i-th element in the j-th row, I c is the number of categories.
6. The multi-user satellite-to-ground communication method based on location semantic information according to claim 1, characterized in that: The method of combining the semantic feature matrix M with the prior channel knowledge vector to perform channel coding and generate the signal matrix X includes: The semantic feature matrix M and the prior channel knowledge vector D are embedded and concatenated to generate discrete features. Where, Embedding(·) represents the embedding and concatenation operation, and D is the vector formed by concatenating the LEO satellite velocity and signal-to-noise ratio; According to discrete features Get the signal matrix X: Among them, Dense(·) represents dense block operation, Sigmoid(·) represents the sigmoid function, W3 and b3 are trainable weights and biases, respectively.
7. The communication system of the multi-user satellite-to-ground communication method based on location semantic information according to any one of claims 1 to 6, characterized in that: It includes a location-aware semantic encoder, a knowledge-assisted channel encoder, a knowledge-assisted channel decoder, a location-aware semantic decoder and a multi-user discriminator; A location-aware semantic encoder and a knowledge-assisted channel encoder are arranged at the satellite end, and a knowledge-assisted channel decoder, a location-aware semantic decoder and a multi-user discriminator are arranged at the user end; A position-aware encoder is used to semantically encode the matrix S based on a multi-head attention mechanism and convert it into a semantic feature matrix M. The multi-head attention mechanism constructs an attention score based on the relative position information between words. The knowledge-assisted encoder is used to combine the semantic feature matrix M with the prior channel knowledge vector to perform channel coding and generate the signal matrix X; After the signal matrix X is transmitted via the satellite-ground channel, the signal Y received by user k is k , k = 1, 2…K, K represents the number of users; A knowledge-assisted decoder is used to convert the received signal Y k Restore to semantic feature matrix The position-aware decoder is used to transform the semantic feature matrix into Convert to sentence matrix Multi-user discriminator for sentence matrix Separate and obtain the sentence content of the corresponding user.
8. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the multi-user satellite-to-ground communication method based on location semantic information are implemented as claimed in any one of claims 1 to 6.
9. A multi-user satellite-to-ground communication device based on location semantic information, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that: The processor executes the computer program to implement the steps of the multi-user satellite-to-ground communication method based on location semantic information according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the multi-user satellite-to-ground communication method based on position semantic information as claimed in any one of claims 1 to 6 are implemented.