A semantic communication transmission method based on non-orthogonal multiple access

Abstract

The present invention discloses a semantic communication transmission method based on non-orthogonal multiple access. The base station constructs a semantic background knowledge base based on a text dataset and shares it with multiple users in the network via broadcast. Transformer-based transmitters and receivers are designed on the base station and user sides, respectively, to extract semantic information on the user side and recover it on the base station side. A non-orthogonal multiple access semantic communication system is designed, which uses semantics to assist in performing serial interference cancellation. The user's transmitted signals and the base station's received signals are collected, and a network model is trained using cross-entropy loss. Based on the trained model parameters, a greedy algorithm and a semantic decoder are combined to recover user data and achieve semantic information transmission. This method uses semantic information to assist in interference cancellation, accurately reconstructs user signals, reduces error propagation, and updates the user decoder network parameters on the base station side, ensuring transmission accuracy for each user while also reducing the base station's processing load and consumption.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a semantic communication transmission method based on non-orthogonal multiple access. Background Art

[0002] For the past few decades, communications have primarily focused on accurately transmitting signals measured in bits. In such systems, the bit error rate (BER) is the primary measure of system performance. With the continuous advancement of communications technology, current transmission rates have increased tens of thousands of times, and system capacity is gradually approaching the Shannon limit. Traditional communication models are facing significant challenges and struggle to meet the demands of many emerging applications. Semantic communication is gradually gaining attention, hoping to leverage its advantages to improve communication efficiency. Semantic communication systems aim to convey the meaning of sentences, rather than the bits transmitted in traditional communication. Only information that is important, relevant, and useful to users can be extracted and transmitted from large amounts of data. This approach significantly reduces data traffic and significantly improves transmission efficiency and reliability. Such systems are particularly useful in typical communication systems with limited bandwidth and low signal-to-noise ratios (SNRs).

[0003] Designing a suitable multiple access technology is one of the most important ways to improve system capacity. Multiple access technologies can be roughly divided into two different methods, namely orthogonal multiple access (OMA) and non-orthogonal multiple access (NOMA). NOMA can use a single frequency resource to serve multiple users at the same time and is a strong candidate for the next generation of wireless access networks. Compared with the well-known high-capacity OMA technology orthogonal frequency division multiple access (OFDMA), NOMA can allocate a frequency channel to multiple users simultaneously in the same cell and has many advantages, including improved spectral efficiency (SE), higher cell edge throughput, relaxed channel feedback, and low transmission delay. Current research on NOMA focuses on aspects such as power allocation and system capacity, mainly by considering the Shannon capacity theorem to solve application problems.

[0004] Integrating the concept of semantic communication with non-orthogonal multiple access systems can further improve spectral efficiency. Currently, research on semantic communication systems primarily focuses on traditional communication frameworks, without integrating them with NOMA. Currently, most semantic communication research has remained within the framework of end-to-end communication, failing to fully leverage the advantages of semantic transmission. Unlike traditional communication methods that involve signal modulation and demodulation, the encoding and decoding structures of semantic communication are inconsistent, preventing simple docking of input and output. Furthermore, the order in which data is processed differs. Summary of the Invention

[0005] Purpose of the Invention: The present invention provides a semantic communication transmission method based on non-orthogonal multiple access. This method can implement non-orthogonal multiple access system transmission within a semantic communication framework, introduce semantic information to assist in serial interference cancellation, and implement user information decoding, thereby improving spectrum efficiency. This overcomes the problem of extensive information redundancy in traditional communications and optimizes the timing of base station signal detection and user data decoding.

[0006] Technical solution: The semantic communication transmission method based on non-orthogonal multiple access of the present invention comprises the following steps:

[0007] S1. A semantic background knowledge base is built at the base station based on the text dataset, and the semantic background knowledge base is shared with each user in the network through broadcasting. The user uses the semantic background knowledge base to transcode the sent information;

[0008] S2. Design a Transformer-based transmitter and receiver on the base station side and the user side, respectively, to extract and compress semantic information on the user side and restore semantic information on the base station side.

[0009] S3. Build an uplink non-orthogonal multiple access semantic communication system, retain semantic information to achieve signal reconstruction, and assist in performing serial interference cancellation;

[0010] S4. Design a unified decoding network for user signals on the base station side to obtain received data, collect user sent data, calculate the cross-entropy loss function, and train the deep network model of the transceiver side of the non-orthogonal multiple access semantic communication system;

[0011] S5. Based on the trained model parameters, the greedy algorithm and semantic decoder are combined to recover the user data received by the base station and realize semantic information transmission.

[0012] Furthermore, in step S1, a semantic background knowledge base is constructed at the base station based on the text data set, and is shared with multiple users in the network through broadcasting. The users use the semantic knowledge base to transcode the information to be sent, specifically: the text data set is preprocessed as a training set, the sentence length is divided into 4 to 30 words, the character format is standardized, and a corresponding dictionary is generated according to the number of words, i.e., the semantic background knowledge base, which is used to realize the mutual conversion between sentence information and signal stream. In order for the transceiver to understand the semantic information, the base station broadcasts the semantic knowledge base to each user in the network, and the user sending end transcodes the user information according to the knowledge base, and converts the sent sentence into a complex signal stream. The signals of each user do not affect each other and follow consistent transcoding rules. The data sent by user i is recorded as a set of sentences S i =[s1,s2,…,s N ]. Among them, a single sentence s n =[w1,w2,…,w L],n=[1,2,…,N],w l represents the lth word in the sentence. User signals are superimposed across their respective channels and sent to the base station. Upon receiving the superimposed user signals, the base station performs serial interference cancellation to detect the signals of different users. After decoding and recovery, the base station recovers each user's original data based on the knowledge base and restores the user data into a complete sentence using the dictionary index.

[0013] Furthermore, in step S2, a transmitter and receiver based on the Transformer structure are designed on the base station and user sides respectively to realize semantic information extraction and compression on the user side and semantic information recovery on the base station side. Specifically, the semantic encoder and decoder are composed of 4 Transformer encoder layers and 4 Transformer decoder layers respectively. At the user sending end, the user data is first encoded by the semantic encoder to extract and compress the semantic information, and then channel coding is performed to ensure the successful transmission of the semantic information on the physical channel. The signal sent by the user can be expressed as: X i =C α (E β (S)), where C α represents the channel encoder with parameter α, E β Represents a semantic encoder network with parameter β. After the base station receives the signal and completes signal detection and serial interference cancellation, it decodes and recovers the signal of each user separately. The receiver includes a signal reconstruction network and a decoding recovery network, where the signal reconstruction network reconstructs the user signal and removes it from the superimposed signal when performing serial interference cancellation. In the signal reconstruction network, we divide the reconstruction methods into three types: channel reconstruction, semantic reconstruction, and direct reconstruction. Among them, when performing semantic reconstruction, after the base station completes decoding the user signal with strong received power, it is necessary to introduce a Transformer-based semantic encoder. Since the signal dimension changes after decoding, we remove the embedding layer in the encoder and meet the re-encoding requirements by supplementing constants of specific dimensions. In the decoding recovery network, after the serial interference cancellation is completed, the base station detects the signals from different users, recovers the transmitted signal through channel decoding, and then combines the background knowledge base to restore the signal to the data sent by the user through the semantic decoder. The recovered data is represented as: in represents the channel decoder with parameter δ, represents the semantic decoder with parameter χ.

[0014] Furthermore, in step S3, an uplink non-orthogonal multiple access semantic communication system is constructed and designed to retain semantic information for signal reconstruction and assist in serial interference cancellation. Specifically, a semantic communication scheme based on uplink non-orthogonal multiple access is considered, where one base station serves two users. Both the user and the base station are equipped with single antennas. To ensure the difference in received power at the base station, the user channel gain state |h1| is assumed. 2 >|h2| 2 , the two users are recorded as the near user and the far user respectively. The users send signals to the base station with their respective transmission powers. The transmission power of the two users P i The base station receives the signal from the two users, which is the superposition of the two user signals. After signal detection, the semantic information is retained, and the signal is reconstructed with the help of the channel codec, channel interference is eliminated, and serial interference cancellation is performed to recover the user data. The base station performs serial interference cancellation as follows: After receiving the superposition signal from the two users, the signals of different users are detected and decoded through serial interference cancellation. The signal transmitted by user i is denoted as X. i , the signal received by the base station is Among them H i is the Rayleigh fading channel between the user and the base station, Z is the variance The user signal detection is realized by the minimum mean square error algorithm, and the formula is:

[0015]

[0016] First, the signal of the near user is detected, and the signal of the far user is regarded as interference. The semantic information is introduced to assist, eliminate the channel influence and detect the near user signal. Reconstruction is performed. Specifically, after detecting the near user signal using the minimum mean square error, the signal needs to be processed to eliminate transmission errors and channel interference and assist in the detection of the far user signal. Define the processing network as R(·), then the processed signal is expressed as In traditional communications, when reconstructing the received signal, it is necessary to first demodulate the signal, then perform channel decoding, and then re-encode the modulation. Based on this idea, a semantic reconstruction scheme is proposed. The modulation in the traditional scheme is converted into semantic coding under the semantic communication framework. After the channel decoding of the near-user signal is completed, semantic decoding is performed instead of the demodulation in traditional communication, and then re-encoding is performed. Since the dimension of the signal changes after semantic decoding, it is necessary to add a constant of a specific dimension to meet the coding requirements, and reduce the coding error through training. During re-encoding, the embedding layer in the semantic encoder is removed to ensure that the recovered signal is consistent with the transmitted signal. However, in the semantic communication framework, if the semantic information is decoded and re-extracted, the compressed semantic information extracted from the current signal will be lost. On this basis, a channel reconstruction scheme is further proposed, that is, retaining the semantic information in the signal, and reducing the impact of channel interference and signal reconstruction through channel decoding. In order to reflect the accuracy of channel reconstruction, we also use the direct reconstruction scheme as a comparison scheme, that is, without any signal processing, the detected signal is directly used for reconstruction, that is, We assume that the receiver has perfect channel state information, and the reconstruction method is:

[0017]

[0018] After the reconstruction of the near user signal is completed, the user signal is removed from the total received signal, that is, Y2 = Y-Y1. After removing the interference of the near user, the remaining signal Y2 only contains the far user signal and the Gaussian white noise interference Z in the channel. The signal of the far user is detected from the signal Y2 to obtain Therefore, the signals of different users are separated. Therefore, compared with far users, near users are subject to greater interference, and the accuracy of near user decoding and reconstruction has a greater impact on far users.

[0019] Furthermore, in step S4, the user decoding network parameters are jointly trained, the transmitted and received signals of the communication are collected, and the network model is trained based on the cross-entropy loss. Specifically, during the local training phase, the transmitted and received signals of each user are collected, and the cross-entropy loss is used as the optimization function to obtain the transmission loss values ​​for the two users respectively. The deep network parameters of the transceiver of the non-orthogonal multiple access semantic communication system are updated in real time based on the average signal loss. The base station is equipped with a receiver to decode and recover the signals of the two detected users respectively. For the signals received from different users, the base station assigns the same signal decoding network structure. If these decoding networks are updated independently, that is, trained independently for different user signals, the distant user will be severely interfered with, unable to transmit the signal correctly, and the average loss value of the two users will be difficult to correctly reduce. Therefore, we uniformly train the parameters of the different decoding networks to ensure that the data of each user can be correctly transmitted.

[0020] Furthermore, in step S5, based on the trained model parameters, the greedy algorithm and the semantic decoder are combined to restore the user data and realize the transmission of semantic information. Specifically, in the training stage, the signal recovered by the base station is used for the cross entropy loss function, and the training result is obtained by calculating the difference between the predicted probability distribution and the true probability distribution. It cannot be directly restored to the sentence corresponding to the background semantic knowledge base. In practical applications, the trained semantic decoder network is combined with the greedy algorithm to take the result with the maximum probability at each word position in the decoded recovery signal, discard the remaining results and reorganize them into a symbol stream. Combined with the semantic knowledge base, the base station can convert the symbol stream back to the text information that the user wants to send. The text recovered by the receiving end is compared with the original text sent by the user to obtain the BLEU score and semantic similarity to evaluate the accuracy of the recovered sentence. Among them, for a sentence s of length l sent, the receiving end recovers the length Sentences The BLEU score is calculated as:

[0021]

[0022] where u n is the weight of n-grams, and p n is the score of n-grams, expressed as: C k (·) is the frequency count function of the kth element in the nth gram. Since the BLEU score cannot understand semantic information and can only judge the transmission accuracy based on sentence length and word accuracy, semantic similarity is added as a scoring criterion at the semantic level, and its calculation method is:

[0023]

[0024] Among them B φ (·) is a pre-trained BERT model containing billions of parameters for extracting semantic meaning. The same process is used to restore sentence information for each user's signal. Furthermore, using the trained decoder model, the signal can be directly restored to a symbol stream when reconstructing the near-user signal. In this case, using the same semantic encoder as the transmitter can meet the signal reconstruction requirements without further consideration of changes in signal dimensionality.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0026] (1) This semantic information-assisted serial interference elimination method for non-orthogonal multiple access systems, combined with a semantic communication framework, breaks through the limitation of insufficient spectrum resources in traditional schemes by extracting useful semantic information and filtering out irrelevant redundant information. At the same time, combined with the non-orthogonal multiple access system, it further improves the spectrum efficiency.

[0027] (2) This method of serial interference elimination in non-orthogonal multiple access systems based on semantic information assistance compares and analyzes three types of signal reconstruction methods, and proposes to use semantic information to assist in interference elimination during serial interference elimination, retain the semantic information extracted from the signal, achieve accurate reconstruction of user signals, reduce error propagation, and reduce interference to other users.

[0028] (3) This method of serial interference elimination in non-orthogonal multiple access systems based on semantic information assistance proposes to jointly train the decoder network parameters of different users and assist training between users, which can reduce the interference of near users on far users and ensure the transmission accuracy of the two users. At the same time, it can also reduce the processing volume and consumption during training and reduce the processing complexity of the receiver. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention is a flowchart of a method for eliminating serial interference in a non-orthogonal multiple access system based on semantic information assistance according to an embodiment of the present invention.

[0030] Figure 2 This is a diagram of an uplink non-orthogonal multiple access communication network architecture designed in an embodiment in which a single base station serves two users.

[0031] Figure 3 Schematic diagram of the network structure of three user signal reconstruction solutions proposed in the embodiment.

[0032] Figure 4 is a graph of average BLUE scores of users under the condition of jointly designed user receiver parameters in the embodiment;

[0033] Figure 5 is a graph of average semantic similarity scores under the condition of jointly designing user receiver parameters in the embodiment;

[0034] Figure 6 is the average fraction of user data when jointly designing user receivers and separately designing receiver parameters in the embodiment, wherein the reconstruction method of the near user is channel reconstruction. DETAILED DESCRIPTION

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] Example

[0037] A semantic communication transmission method based on non-orthogonal multiple access, such as Figure 1 , including the following steps,

[0038] S1. A semantic background knowledge base is built at the base station based on the text dataset and shared with multiple users in the network through broadcasting. Users use the semantic knowledge base to transcode the information they send.

[0039] In order to simulate and verify the actual application effect of this method, the embodiment processes the text training set and counts the word frequencies, generates a dictionary and provides it to different user transmitters and base station receivers as a semantic background knowledge base. According to the knowledge base, the transmitter transcodes the user information, converts the sentence form into a complex symbol stream, and sends it to the base station receiver through a Rayleigh fading channel after encoding. The base station receives the user signal and decodes it to restore it to a symbol stream, and then combines the background knowledge base to restore the symbol to user information. The text data set is preprocessed as a training set, the sentence length is divided into 4 to 30 words, the character format is standardized, and the corresponding dictionary is generated according to the number of words, that is, the semantic background knowledge base, which is used to realize the mutual conversion between sentence information and signal stream. In order for the transceiver to understand the semantic information, the base station broadcasts the semantic knowledge base to each user, and the user transmitter transcodes the user information according to the knowledge base and converts the sent sentence into a complex signal stream. The signals of each user do not affect each other and follow consistent transcoding rules. The data sent by user i is recorded as a set of sentences S i =[s1,s2,…,s N ]. Among them, a single sentence s n =[w1,w2,…,w L ],n=[1,2,…,N],w l represents the lth word in the sentence. User signals are superimposed across their respective channels and sent to the base station. Upon receiving the superimposed user signals, the base station performs serial interference cancellation to detect the signals of different users. After decoding and recovery, the base station recovers each user's original data based on the knowledge base and restores the user data into a complete sentence using the dictionary index.

[0040] S2. Design a Transformer-based transmitter and receiver on the base station and user side, respectively, to extract and compress semantic information on the user side and restore it on the base station side.

[0041] In order to implement the semantic communication framework, it is necessary to extract the semantic information in the user data and compress it to meet the semantic communication requirements. In the embodiment, the semantic encoder and decoder are composed of 4 Transformer encoder layers and 4 Transformer decoder layers respectively. At the user sending end, the user data is first encoded by the semantic encoder to extract and compress the semantic information, and then channel coding is performed to ensure the successful transmission of the semantic information on the physical channel. The signal sent by the user can be expressed as: X i =C α (E β (S)), where C α represents the channel encoder with parameter α, E βRepresents a semantic encoder network with parameter β. After the base station receives the signal and completes signal detection and serial interference cancellation, it decodes and recovers the signal of each user separately. The receiver includes a signal reconstruction network and a decoding network, wherein the signal reconstruction network reconstructs the user signal and removes it from the superimposed signal when performing serial interference cancellation. In the signal reconstruction network, we divide the reconstruction methods into three types: channel reconstruction, semantic reconstruction, and direct reconstruction. Among them, when performing semantic reconstruction, after the base station completes decoding of the user signal with strong received power, it is necessary to introduce a Transformer-based semantic encoder. Since the signal dimension changes after decoding, we remove the embedding layer in the encoder and meet the re-encoding requirements by supplementing constants of specific dimensions. After the serial interference cancellation is completed, the base station detects the signals from different users, recovers the transmitted signal through channel decoding, and then combines the background knowledge base to restore the signal to the data sent by the user through the semantic decoder. The recovered data is represented as: in denotes the channel decoder with parameters δ, and denotes the semantic decoder with parameters χ.

[0042] S3. Build and design an uplink non-orthogonal multiple access semantic communication system to preserve semantic information for signal reconstruction and assist in serial interference cancellation.

[0043] like Figure 2 As shown, to achieve user communication in a non-orthogonal multiple access system, the embodiment designs a communication network in which a single base station serves two users. Both the base station and the user end are equipped with a single antenna. To ensure that the received power of different users at the base station is different, the user channel gain state |h1| 2 >|h2| 2 , the two users are recorded as the near user and the far user respectively. The users send signals to the base station with their respective transmission powers. The transmission power of the two users P i The signals received by the base station are the superposition of the two user signals and P1=P2=1. After signal detection, the semantic information is retained and the signal is reconstructed with the help of the channel codec to eliminate channel interference and assist in performing serial interference cancellation, thereby ensuring that the user data can be accurately recovered.

[0044] The process of serial interference cancellation at the base station receiving end is as follows: at the receiving end, after the base station receives the superimposed signals from two users, it detects the signals from different users through serial interference cancellation and decodes them. The signal transmitted by user i is denoted as X i , the signal received by the base station is Among them H i is the Rayleigh fading channel between the user and the base station, Z is the variance Independent zero-mean additive Gaussian white noise. The minimum mean square error algorithm is used to detect user signals. The formula is: ( is the total noise variance of user i. For near users, it includes the interference caused by far users and the channel noise interference. For far users, since the near user signal has been removed from the total received signal during signal detection, only the channel noise interference exists.

[0045]

[0046] First, the signal of the near user is detected, and the signal of the far user is regarded as interference. The semantic information is introduced to assist, eliminate the channel influence and detect the near user signal. Reconstruction is performed. Specifically, after detecting the near user signal using the minimum mean square error, the signal needs to be processed to eliminate transmission errors and channel interference and assist in the detection of the far user signal. Define the processing network as R(·), then the processed signal is expressed as In traditional communications, when reconstructing the received signal, it is necessary to first demodulate the signal, then perform channel decoding, and then re-encode the modulation. Based on this idea, a semantic reconstruction scheme is proposed. The modulation in the traditional scheme is converted into semantic coding under the semantic communication framework. After the channel decoding of the near-user signal is completed, semantic decoding is performed instead of the demodulation in traditional communication, and then re-encoding is performed. Since the dimension of the signal changes after semantic decoding, it is necessary to add a constant of a specific dimension to meet the coding requirements, and reduce the coding error through training. During re-encoding, the embedding layer in the semantic encoder is removed to ensure that the recovered signal is consistent with the transmitted signal. However, in the semantic communication framework, if the semantic information is decoded and re-extracted, the compressed semantic information extracted from the current signal will be lost. On this basis, a channel reconstruction scheme is further proposed, that is, retaining the semantic information in the signal, and reducing the impact of channel interference and signal reconstruction through channel decoding. In order to reflect the accuracy of channel reconstruction, we also use the direct reconstruction scheme as a comparison scheme, that is, without any signal processing, the detected signal is directly used for reconstruction, that is, We assume that the receiver has perfect channel state information, and the reconstruction method is:

[0047]

[0048] After reconstructing the near-user signal, the signal is removed from the total received signal, i.e., Y2 = Y - Y1. After removing the near-user interference, the remaining signal contains only the far-user signal and the Gaussian white noise interference in the channel. Minimum mean square error detection is then performed on the far-user signal within the remaining signal to separate the signals of different users. Therefore, near-users experience greater interference than far-users, and the accuracy of near-user decoding reconstruction significantly affects the accuracy of far-user decoding.

[0049] During the signal detection process, the accuracy of user signal reconstruction plays a key role in the signal transmission of all users in the system. This embodiment proposes a channel reconstruction scheme, which uses the minimum mean square error algorithm as the signal detection algorithm, and performs channel decoding on the detected near-user signal to minimize the interference caused by the physical channel. The processed signal contains the encoded semantic information sent by the user. On the premise of retaining the encoded information, repeated channel coding is performed to restore the signal state before the transmitter enters the physical channel to achieve signal reconstruction. This scheme can maximize the retention of the semantic information sent and encoded by the user while removing the physical channel noise, and has a low impact on subsequent semantic decoding. If the semantic information in the signal is reconstructed after channel decoding, that is, the semantic reconstruction scheme, the semantic information will suffer a certain loss, which will have a greater impact on the accuracy of signal transmission. If the transmitted user signal is not reconstructed, that is, the signal after passing through the physical channel is directly used as the reconstructed signal, the user information cannot be correctly restored, and the signal transmission is greatly affected. The network structures of the three types of reconstruction schemes are as follows: Figure 3 As shown in Figure 2, the performance comparison of the three schemes, channel reconstruction, semantic reconstruction, and direct reconstruction, is shown in Figure 2. Using semantic similarity and BLEU scores as the judgment benchmarks, respectively. Figure 4 and Figure 5 shown.

[0050] S4, jointly train the user decoding network parameters, collect the communication transmission and reception signals, and train the network model based on the cross entropy loss;

[0051] During the local training phase, the transmission signal and the received signal of each user are collected, and the cross entropy loss is used as the optimization function to obtain the transmission loss values ​​of the two users respectively, and the deep network parameters of the transceiver of the non-orthogonal multiple access semantic communication system are updated in real time according to the mean value of the signal loss. The base station is equipped with a receiver to decode and recover the two detected user signals respectively. For different user signals received, the base station allocates the same signal decoding network structure. According to the semantic background knowledge base shared by the transmitter and the receiver, that is, the text data set, the user's transmission signal is encoded, the semantic information is extracted and the symbol is converted into a complex signal stream, and then the encoded signal is transmitted to the base station receiver through the physical channel. At the base station end, when decoding the user signal, the embodiment first decodes and recovers the near user signal, and then decodes the far user signal from the remaining signal. The embodiment adopts channel reconstruction to realize the recovery of the near user signal, and jointly designs and updates the decoder parameters of different users, so as to avoid the interference caused by the near user to the far user during signal decoding during the training process. As Figure 6Shown is a comparison of the average scores of the transmission accuracy of the two users when the parameters are jointly updated and when the parameters are updated independently. It can be concluded that, during training, the joint design of the decoding parameters of the two users can effectively ensure the reception accuracy. When the decoding parameters of the two user signals are different, the average accuracy of the users is low. When training the network, the embodiment uses the same decoding network parameters for the signals of different users to decode different users, and uniformly optimizes them according to the training loss, thereby avoiding the problem that the network parameters of distant users are difficult to update correctly during decoding. In the case of independent updating of model parameters, the update of the decoding model parameters of distant users is interfered with by nearby users, resulting in the inability to correctly update the model during training, and the decoding signal of the distant user suffers from large loss, making it impossible to transmit data correctly.

[0052] S5. Based on the trained model parameters, the greedy algorithm and semantic decoder are combined to restore user data and realize semantic information transmission.

[0053] During the training phase, the signal recovered by the base station is applied to a cross-entropy loss function. The training result is calculated by calculating the difference between the predicted probability distribution and the true probability distribution. This does not directly correspond to the background semantic knowledge base to recover the sentence. In an embodiment, after training is complete, the base station receiver uses a greedy algorithm combined with the trained semantic decoder parameters to decode the user signal. The semantic information sent by the user is then recovered using the background semantic knowledge base, thereby enabling signal transmission. After eliminating channel interference, the base station performs greedy decoding to extract the semantic information from the signal and reconstruct it into a sentence. The decoded result with the highest probability at each word position in the recovered signal is taken, and the remaining results are discarded to reconstruct the symbol stream. Combined with the semantic knowledge base, the base station can convert the symbol stream back into the text message the user intended to send. The text recovered by the receiver is compared with the original text sent by the user. In this embodiment, semantic similarity and the Bilingual Evaluation Undertaking (BLEU) score are used to assess the accuracy of the recovered semantic information. The BLEU score is a commonly used evaluation metric in natural language processing. It can compare and judge the accuracy of the target sentence's words, but it cannot understand semantic information and cannot be combined with semantic scoring. Therefore, semantic similarity is used as another evaluation metric to score the accuracy of the transmitted sentence at the semantic level. For a sent sentence s of length l, the receiver recovers the length Sentences The BLEU score is calculated as:

[0054]

[0055] where u n is the weight of n-grams, and p n is the score of n-grams, expressed as: C k (·) is the frequency counting function of the kth element in the nth grams.

[0056] Semantic similarity is a scoring criterion at the semantic level and is calculated as follows:

[0057]

[0058] Among them B φ (·) is a pre-trained BERT model containing billions of parameters for extracting semantic meaning. The same process is used to restore sentence information for each user's signal. Furthermore, using the trained decoder model, the signal can be directly restored to a symbol stream when reconstructing the near-user signal. In this case, using the same semantic encoder as the transmitter can meet the signal reconstruction requirements without further consideration of changes in signal dimensionality.

[0059] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A semantic communication transmission method based on non-orthogonal multiple access, characterized in that: The steps include: S1. A semantic background knowledge base is built at the base station based on the text dataset, and the semantic background knowledge base is shared with each user in the network through broadcasting. The user uses the semantic background knowledge base to transcode the sent information; S2. Design a Transformer-based transmitter and receiver on the base station side and the user side, respectively, to extract and compress semantic information on the user side and restore semantic information on the base station side. S3. Build an uplink non-orthogonal multiple access semantic communication system, retain semantic information to achieve signal reconstruction, and assist in performing serial interference cancellation; Step S3 is specifically as follows: Consider a semantic communication scheme based on uplink non-orthogonal multiple access, where one base station serves two users. Both the user and the base station are equipped with single antennas. Let the user channel gain state |h1| 2 >|h2| 2 , the two users are recorded as the near user and the far user respectively; the users send signals to the base station with their own transmission power, and the transmission power of the two users P i If the two user signals are the same and P1=P2=1, the signal received by the base station is the superposition of the two user signals. After signal detection, the semantic information is retained and the signal is reconstructed with the help of the channel codec, channel interference is eliminated, and serial interference cancellation is performed to recover the user data. The base station implements serial interference cancellation as follows: after receiving the superimposed signals from two users, it detects the signals of different users and decodes them through serial interference cancellation; the signal transmitted by user i is X i , the signal received by the base station is Among them H i is the Rayleigh fading channel between the user and the base station, Z is the variance Independent zero-mean additive Gaussian white noise; the minimum mean square error algorithm is used to detect user signals, the formula is: First, the signal of the near user is detected, and the signal of the far user is regarded as interference. The semantic information is introduced to assist, eliminate the channel influence and detect the near user signal. After reconstructing and detecting the near user signal using the minimum mean square error, the signal is processed to eliminate transmission errors and channel interference, and assist in the signal detection of far users. The processing network is defined as R(·), and the processed signal is expressed as After the near-user signal channel decoding is completed, semantic decoding is performed and then re-encoding is performed; a constant of dimension is added to meet the encoding requirements, and the encoding error is reduced through training; during re-encoding, the embedding layer in the semantic encoder is removed to ensure that the recovered signal is consistent with the transmitted signal; the detected signal is used for reconstruction, that is, Assuming that the receiving end has channel state information, the reconstruction method is: After the reconstruction of the near user signal is completed, the user signal is removed from the total received signal, that is, Y2 = Y-Y1; after removing the interference of the near user, the remaining signal Y2 only contains the far user signal and the Gaussian white noise interference Z in the channel. The signal of the far user is detected from the signal Y2 to obtain Thus, the signals of different users are separated; S4. Design a unified decoding network for user signals on the base station side to obtain received data, collect user sent data, calculate the cross-entropy loss function, and train the deep network model of the transceiver side of the non-orthogonal multiple access semantic communication system; Step S4 specifically includes: in the local training phase, collecting the transmitted and received signals of each user, using the cross entropy loss as the optimization function, obtaining the transmission loss values ​​of the two users respectively, and updating the deep network parameters of the transceiver of the non-orthogonal multiple access semantic communication system in real time according to the mean signal loss value; The base station is equipped with a receiver to decode and recover the signals of the two detected users. For each received user signal, the base station assigns the same signal decoding network structure and uniformly trains the parameters of different decoding networks to ensure that each user's data can be correctly transmitted. S5. Based on the model parameters after deep network model training, the greedy algorithm and semantic decoder are combined to restore the user data received by the base station and realize semantic information transmission.

2. The semantic communication transmission method based on non-orthogonal multiple access according to claim 1, characterized in that: Step S1 is specifically as follows: pre-processing the text dataset as a training set, dividing the sentence length into 4 to 30 words, standardizing the character format and generating a corresponding dictionary according to the number of words, i.e., a semantic background knowledge base, for realizing the mutual conversion between sentence information and signal stream; in order to enable the transceiver to understand the semantic information, the base station broadcasts the semantic knowledge base to each user in the network, and the user sending end transcodes the user information according to the knowledge base and converts the sent sentence into a complex signal stream, where the signals of each user are independent of each other and follow the same transcoding rules. The data sent by user i is a set of sentences: S i =[s1,s2,…,s N ], where a single sentence s n =[w1,w2,…,w L ],n=[1,2,…,N],w l Represents the lth word in the sentence; user signals are superimposed through their respective channels and sent to the base station; after receiving the superimposed user signals, the base station performs serial interference cancellation to detect the signals of different users, and after decoding and recovery, restores the original data of each user according to the knowledge base, and restores the user data into a complete sentence according to the dictionary index.

3. The semantic communication transmission method based on non-orthogonal multiple access according to claim 1, characterized in that: Step S2 is as follows: the semantic encoder and decoder are composed of 4 Transformer encoder layers and 4 Transformer decoder layers respectively; at the user sending end, the user data is first encoded by the semantic encoder to extract and compress the semantic information, and then channel coding is performed to ensure the successful transmission of the semantic information on the physical channel. The signal sent by the user is expressed as: X i =C α (E β (S)), where C α represents the channel encoder with parameter α, E β The semantic encoder network is represented by parameter β. After the base station receives the signal and completes signal detection and serial interference cancellation, it decodes and recovers the signal of each user separately. The receiver includes a signal reconstruction network and a decoding recovery network. The signal reconstruction network reconstructs the user signal and removes it from the superimposed signal when performing serial interference cancellation. In the signal reconstruction network, the reconstruction methods are divided into three types: channel reconstruction, semantic reconstruction, and direct reconstruction. When performing semantic reconstruction, after the base station completes decoding of the user signal with strong received power, it introduces a Transformer-based semantic encoder. The embedding layer in the encoder is removed, and the dimensional constant is added to meet the re-encoding requirements. In the decoding recovery network, after completing serial interference cancellation, the base station detects the signals from different users, recovers the transmitted signal through channel decoding, and then recovers the signal into the data sent by the user through the semantic decoder in combination with the background knowledge base. The recovered data is represented as: in represents the channel decoder with parameter δ, represents the semantic decoder with parameter χ.

4. The semantic communication transmission method based on non-orthogonal multiple access according to claim 1, characterized in that: Step S5 is specifically as follows: in the training phase, the signal recovered by the base station is used for the cross entropy loss function, and the training result is obtained by calculating the difference between the predicted probability distribution and the true probability distribution. The trained semantic decoder network is combined with the greedy algorithm to obtain the result with the maximum probability at each word position in the decoded recovery signal, discard the remaining results and reorganize them into a symbol stream. Combined with the semantic knowledge base, the base station converts the symbol stream back to the text information that the user wants to send, and compares the text recovered by the receiving end with the original text sent by the user to obtain the BLEU score and semantic similarity for evaluating the accuracy of the recovered sentence; wherein, for a sentence s of length l sent, the receiving end recovers the length Sentences The BLEU score is calculated as: where u n is the weight of n-grams, and p n is the score of n-grams, expressed as: C k (·) is the frequency counting function of the kth element in the nth grams. Semantic similarity is added as the scoring criterion at the semantic level, which is calculated as follows: Among them B φ (·) is a pre-trained BERT model containing billions of parameter information for extracting semantics. The same process is used to restore sentence information for each user's signal. The trained decoder model is used to directly restore the signal to a symbol stream when reconstructing the near-user signal. The semantic encoder consistent with the sender is used to meet the signal reconstruction requirements.

Images