Semantic adaptation layer training method based on transmitting terminal model

Abstract

The invention provides a semantic adaptation layer training method based on a transmitting terminal model, which relates to the technical field of intelligent communication, and comprises the following steps: for unmatched users, a base station constructs a semantic adaptation layer; for each unmatched user, the base station uses the transmitting end model to encode and transmit the training data, an obtained first receiving signal passes through a semantic adaptation layer to obtain a third receiving signal, and meanwhile, the transmitting end model matched with the receiving end model of the unmatched user is used to encode and transmit the training data to obtain a second receiving signal; taking a mean square error between the second receiving signal and the third receiving signal as a training loss function; after training is completed, the base station sends parameters of all the semantic adaptation layers to the corresponding unmatched users, and all the unmatched users construct and start the semantic adaptation layers of the unmatched users. The semantic communication performance under the condition that the knowledge base is incomplete can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent communication technology, and in particular to a semantic adaptation layer training method based on a transmitter model. Background Technology

[0002] Existing semantic communication systems primarily employ an end-to-end architecture, which can lead to a mismatch between the semantic knowledge bases of the sending and receiving ends, resulting in degraded semantic communication performance. To address this issue, intelligent assisted relays for semantic communication have been proposed. However, intelligent relay-based methods require the use of a shared knowledge base to decode and re-encode the transmitted data, which results in a large number of models required for the relay, making them unsuitable for communication scenarios with limited relay memory or without a relay.

[0003] Furthermore, when using a non-nested update method, the knowledge base mismatch problem in incomplete scenarios is difficult to mitigate effectively through knowledge base recognition schemes, resulting in severe performance degradation and even failure to communicate normally. (See attached...) Figure 2 Taking the broadcast scenario shown as an example, in order to solve the knowledge base mismatch problem under non-nested updates, a simple approach is to have the sender resend the knowledge base to the receiver to overwrite the model it originally stored. However, this solution often requires a large amount of data to be transmitted and waiting for a delay, and it lacks flexibility. Summary of the Invention

[0004] The main objective of this invention is to propose a semantic adaptation layer training method based on the transmitter model. The base station can train the semantic adaptation layer without storing the receiver model of each user, thereby effectively improving the semantic communication performance in the case of incomplete knowledge base.

[0005] This invention is achieved through the following technical solution: A semantic adaptation layer training method based on a transmitter model includes the following steps: Step S1: The base station identifies users whose receiving model does not match the transmitting model used by the base station broadcast as mismatched users, and constructs a semantic adaptation layer. Step S2: The semantic adaptation layer is trained at the base station. The semantic adaptation layer is constructed based on an LSTM network. Step S2: For each mismatched user, the base station encodes the training data using the transmitting model used by the broadcast and transmits it through the constructed channel to obtain the first received signal. The first received signal is then processed by the semantic adaptation layer to obtain the third received signal. At the same time, the base station encodes the training data using the transmitting model that matches the receiving model of the mismatched user and transmits it through the same channel to obtain the second received signal. The mean square difference between the second received signal and the third received signal is used as the loss function for training to update the parameters of the semantic adaptation layer. Step S3, after the training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user, and each mismatched user constructs and enables its semantic adaptation layer according to the received parameters.

[0006] Further, in step S1, all users in communication with the base station first complete access to the base station and inform the base station of the receiver model they have, so that the base station determines the mismatched users.

[0007] Further, the semantic adaptation layer uses a cascaded network of an LSTM network and a full connection layer.

[0008] Further, the transmitter model of the base station includes a DeepSC encoding network with a Transformer encoder, and the receiver model of each user includes a DeepSC decoding network with a Transformer decoder. Further, in step S2, the channel constructed by the base station is an AWGN channel.

[0009] Further, in step S2, the training process of the receiver model and the transmitter model includes nested updating and non-nested updating, the nested updating refers to training the initial model using the training data when the training data is updated, and the non-nested updating refers to training the initial model when the training data is updated.

[0010] Further, in step S2, sentences with a word number between 4 and 30 in the English data set of the European Parliament are used as the training data.

[0011] Further, in step S2, the loss function is represented as , wherein, is a second received signal, is a third received signal, is a semantic adaptation layer parameter, is an average, is a transmitter model matched with the receiver model of the mismatched user B, and the parameter thereof is , is training data, is channel noise.

[0012] From the above description of the present application, compared with the prior art, the present application has the following beneficial effects: The base station of the application constructs a semantic adaptation layer after confirming the mismatched users, and trains the semantic adaptation layer corresponding to each mismatched user respectively. In the training process, the base station uses the transmitter model used in broadcasting and the transmitter model matched with the receiver model of the mismatched user to encode and transmit the training data. The mean square error between the second received signal and the third received signal obtained from the two transmission lines is used as the loss function of the training. After the training is completed, the base station packs the parameters of each trained semantic adaptation layer and sends them to the corresponding mismatched user. Each mismatched user constructs and enables its semantic adaptation layer according to the received parameters. The addition of the semantic adaptation layer can realize semantic communication in the case of incomplete knowledge base. In the training process of the semantic adaptation layer, the receiver model is not needed, so the base station does not need to occupy more memory due to the additional storage of the receiver model of each user, avoiding the adverse effects on the communication performance caused by memory occupation, thereby effectively improving the semantic communication performance. Moreover, the parameter amount of the semantic adaptation layer is much smaller than that of the DeepSC network, so when the parameters of the trained semantic adaptation layer are sent to the mismatched user, the model parameter amount required for transmission can be greatly reduced compared to sending the matched receiver knowledge base to the user, further improving the communication performance. BRIEF DESCRIPTION OF DRAWINGS

[0013] The application will be further described below in combination with the drawings and specific embodiments.

[0014] Figure 1 The flowchart of the application.

[0015] Figure 2 The schematic diagram of the broadcasting scenario of the base station to multiple users.

[0016] Figure 3 The training logic block diagram of the application.

[0017] Figure 4 The communication network architecture based on the application.

[0018] Figure 5 The communication system model based on the application.

[0019] Figure 6 The performance schematic diagram of the application under nested update.

[0020] Figure 7 The performance schematic diagram of the application under non-nested update. DETAILED DESCRIPTION

[0021] The application will be further described below in combination with the drawings and specific embodiments.

[0022] As shown in Figure 1 The semantic adaptation layer training method based on the transmitter model includes the following steps: Step S1, for the mismatched users whose receiver model is different from the sender model used by the base station for broadcasting, the base station constructs a semantic adaptation layer, and enters step S2 to train the semantic adaptation layer, wherein the semantic adaptation layer uses a cascaded network of three LSTM networks and two fully connected layers, the LSTM network has a time memory capability and can well learn the association between words in a sentence, thereby effectively restoring the semantic communication performance; Specifically, all users communicating with the base station first complete access to the base station, and inform the base station of the receiver model number (such as Model A ) they have, and the base station identifies the receiver model of each user according to the number, selects the sender model corresponding to the receiver model with more users for broadcasting, and the remaining users are mismatched users. The base station only stores the receiver model corresponding to each number. Figure 1 A

[0023] In the present application, we define the semantic knowledge base as a semantic encoder and decoder model obtained by training a certain number of rounds based on a certain data set, which respectively correspond to the sender knowledge base (i.e. the sender model) and the receiver knowledge base (i.e. the receiver model). Semantic knowledge base matching means that the training process of the semantic encoder and decoder is completely consistent, that is, the data set and the training rounds are the same and the randomness during training is the same (for example, fixed random seed); semantic knowledge base mismatch means that the training process of the semantic encoder and decoder is not completely consistent.

[0024] When it is desired to allow the base station to perform downlink semantic communication to the user, it is necessary to first extract the characteristic parameters based on the knowledge base of the user and send them to the base station, at which time the base station can identify the model that can be paired with it in the stored model set based on the parameters, that is, the identification of the semantic knowledge base.

[0025] The complete semantic knowledge base scenario is defined as a scenario in which the transmitter and receiver have models that can match each other and can use the models to implement communication, otherwise it is an incomplete semantic knowledge base scenario, and the incomplete semantic knowledge base will certainly lead to the occurrence of semantic knowledge base mismatch. As shown in Figure 2 , the base station performs downlink broadcast semantic communication to two users UE1 and UE2, the receiver model of UE1 user is Model A , and the receiver model of UE2 user is Model B , and the base station can only use one sender model due to the broadcast transmission mode, which will lead to knowledge base mismatch between the base station and part of the users, for example, using the sender Model A transmission, based on the receiver model of UE2 user. In this case, although the transmitter and receiver have models that can match each other, they cannot use the models to implement effective communication, so this scenario is also called an incomplete semantic knowledge base scenario.

[0026] The transmitter model of the base station includes a DeepSC encoding network with a Transformer encoder, and the receiver model of each user includes a DeepSC decoding network with a Transformer decoder. The DeepSC network is a prior art proposed by Xie et al. in the article entitled “Deep learning enabled semantic communication systems” in IEEE Transactions on Signal Processing, vol. 69, 2021.

[0027] In step S2, for each mismatched user, the base station encodes the training data using the transmitter model used for broadcasting and transmits it through the constructed channel to obtain a first received signal, and obtains a third received signal after the semantic adaptation layer. Meanwhile, the base station encodes the training data using the transmitter model matched with the receiver model of the mismatched user and transmits it through the same channel to obtain a second received signal. The mean square error between the second received signal and the third received signal is taken as the loss function of training, and the parameters of the semantic adaptation layer are updated. The training logic block diagram is shown in Figure 3 During training, the channel constructed by the base station is an AWGN channel, and the channel noise is The sentences with the number of words between 4 and 30 in the English data set of the European Parliament are used as training data, and the total number of sentences is 73472. During training, the SNR of the AWGN channel is uniformly selected between 5 and 10 dB, and the number of training rounds is set to 80, the learning rate is , the batch size is 128, the optimizer uses the Adam optimizer, the weight decay is set to , and the betas are set to (0.9, 0.98). The performance test index of the model uses the Bilingual Evaluation Understudy (BLEU) score, and the default calculation unit size is 1.

[0028] In the training of the sender model and the receiver model, the data set is divided into data sets A, B, C and a test set in sequence by using the first 50,000 sentences of the total data set, and the number of sentences contained in the data sets A, B, C and the test set is 20,000, 10,000, 10,000 and 10,000 respectively. The training process of the sender model and the receiver model can be nested updating or non-nested updating. For nested updating, data set A is used for initial training, and at this time, the model based on random initialization is trained, and model A is obtained after training; when the data set is updated (i.e. data set B is supplemented), the model is retrained, but at this time, the model based on random initialization is no longer trained, but based on the previously trained model A. For non-nested updating, data set A is used for initial training, and at this time, the model based on random initialization is trained, and model A is obtained after training; after data set B is supplemented, the model based on initialization is still trained. The specific process of nested updating or non-nested updating is prior art.

[0029] In the training of the semantic adaptation layer, the training set is fixed as the first 40,000 sentences of the total data set, that is, the combination of data sets A, B and C. The vector dimension of the input and output of the semantic adaptation layer is 16. During training, the SNR of the channel is uniformly selected between 5-10 dB, the number of training rounds is also set to 80, the learning rate is , the batch size is 256, the optimizer uses the Adam optimizer, the weight decay is set to 0, and the betas are set to (0.9, 0.999).

[0030] The parameters of the sender model and the receiver model are shown in Table 1, and the parameters of the semantic adaptation layer are shown in Table 2.

[0031] Table 1 Table 2 The loss function used in training is represented as , wherein, is the second received signal, is the third received signal, is the semantic adaptation layer parameter, is the mean value, is the sender model matched with the receiver model of user B, and the parameter of the sender model is , is the training data, is the channel noise, is Model A sender, Model BThe corresponding semantic adaptation layer is received. During the training process, the parameters of the transmitter model and the receiver model are fixed, and the parameters of the semantic adaptation layer are updated .

[0032] Step S3, after the training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user, and each mismatched user constructs and enables its semantic adaptation layer according to the received parameters; The communication network combined with the semantic adaptation layer is constructed as shown in Figure 4 , and the communication system model is as shown in Figure 5 . The base station encodes and broadcasts the communication data to each user. The matched user directly decodes based on its receiver model matched with the transmitter model used in the broadcast, and the mismatched user processes the received signal through the semantic adaptation layer and then decodes based on its receiver model. Among them, Figure 4 , the semantic adaptation layer is trained based on the training method of the present application, is an embedding vector, E is the dimension of the embedding vector corresponding to each word.

[0033] Figure 6 and Figure 7 , Huffman+LDPC is a comparative scheme. For the DeepSC scheme, three model versions are considered, the first, second and third versions are trained using data A, the combination of data A and B, and the combination of data A, B and C, respectively. "DeepSC(A+AB--A)—8 sym / word" and "DeepSC(AB--A)—8 sym / word" represent the performance curves when the first version model is used to communicate with the second version model under nested and non-nested update conditions, respectively; "DeepSC(A+AB+ABC--A)—8 sym / word" and "DeepSC(ABC--A)—8 sym / word" represent the performance curves when the first version model is used to communicate with the third version model under nested and non-nested update conditions, respectively; "DeepSC(A+AB+ABC—A+AB)—8 sym / word" and "DeepSC(ABC--AB)—8 sym / word" represent the performance curves when the second version model is used to communicate with the third version model under nested and non-nested update conditions, respectively. The corresponding curves with the suffix "with TMA" represent the communication performance curves of the corresponding model pair after adding the semantic adaptation layer obtained based on the training scheme of the present application.

[0034] From Figure 6It can be seen that, when the nested updating mode is adopted, the original unmatched model pair can achieve a BLEU performance not lower than that of the conventional scheme under 12.5dB, and meanwhile, the frequency band utilization rate (unit: Baud / Hz) is increased by more than 1.5 times (i.e., the number of symbols required for transmission can be compressed to less than 2 / 3 of that of the conventional scheme under the condition of achieving the same communication effect).

[0035] From Figure 7 It can be seen that, when the non-nested updating mode is adopted, the original unmatched model pair can achieve a BLEU performance not lower than that of the conventional scheme under 12dB, and meanwhile, the frequency band utilization rate is increased by more than 1.5 times.

[0036] In the present application, the terms "first", "second", "third" and the like are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, and cannot be understood as indicating or implying relative importance. In the description, the directions or positional relationships indicated by "up", "down", "left", "right", "front" and "back" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present application, and do not indicate or imply that the devices referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the scope of protection of the present application. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0037] In addition, in the description of the present application, "a plurality of" means two or more, unless otherwise specified. The association relationship between the associated objects described by "and / or" indicates that there can be three relationships, for example, A and / or B can represent the three cases of A alone, A and B together, and B alone. The character " / " generally represents an "or" relationship between the associated objects before and after it.

[0038] The above is only a specific embodiment of the present application, but the design concept of the present application is not limited thereto, and any non-essential modification of the present application using this concept shall be deemed to be an act of infringing the scope of protection of the present application.

Claims

1. A semantic adaptation layer training method based on a transmitter model, characterized in that: The method comprises the following steps: Step S1, the base station determines a user whose receiver model does not match the transmitter model used by the base station for broadcasting as a mismatched user, and constructs a semantic adaptation layer, and enters step S2 to train the semantic adaptation layer at the base station end, wherein the semantic adaptation layer is constructed based on an LSTM network; Step S2, for each mismatched user, the base station encodes training data using the transmitter model used for broadcasting and transmits the training data through the constructed channel to obtain a first received signal, and obtains a third received signal after the first received signal passes through the semantic adaptation layer; the base station simultaneously encodes training data using a transmitter model matched with the receiver model of the mismatched user and transmits the training data through the same channel to obtain a second received signal; the mean square error between the second received signal and the third received signal is taken as a loss function of training, and the parameters of the semantic adaptation layer are updated; Step S3, after the training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user, and each mismatched user constructs and enables its semantic adaptation layer according to the received parameters. 2.The method of training a semantic adaptation layer based on a transmitter model of claim 1, wherein: In the step S1, all users in communication with the base station first complete access to the base station, and inform the base station of their own receiver model, so that the base station determines the mismatched user. 3.The method of training a semantic adaptation layer based on a transmitter model of claim 1, wherein: The semantic adaptation layer uses a cascaded network of an LSTM network and a full connection layer. 4.The training method of a semantic adaptation layer based on a transmitter model according to claim 1 or 2 or 3, characterized in that: Each transmitter model of the base station comprises a DeepSC encoding network with a Transformer encoder, and each receiver model of the user comprises a DeepSC decoding network with a Transformer decoder.

5. The method of claim 1 or 2 or 3, wherein the method further comprises: In the step S2, the channel constructed by the base station is an AWGN channel.

6. The method of claim 1 or 2 or 3, wherein the method further comprises: The training process of the receiver model and the transmitter model comprises nested updating and non-nested updating, the nested updating refers to that when the training data is updated, the model trained previously is trained, and the non-nested updating refers to that when the training data is updated, the initial model is still trained.

7. The method of claim 1 or 2 or 3, wherein the method further comprises: In the step S2, sentences with the number of words between 4 and 30 in the English data set of the European Parliament are taken as the training data. 8.The method of training a semantic adaptation layer based on a transmitter model according to claim 1 or 2 or 3, characterized in that: In the step S2, the loss function is represented as wherein, is the second received signal, is the third received signal, is the semantic adaptation layer parameter, is the mean, is the sender model matched with the receiver model of the non-matching user B, and the parameter of which is , is the training data, is the channel noise.

Images