A semantic adaptation layer training method based on a transmitter model
By constructing a semantic adaptation layer using base stations and training and sending parameters using LSTM networks, the problem of knowledge base mismatch in semantic communication is solved, communication performance and bandwidth utilization are improved, and the number of transmission model parameters is reduced.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2025-11-14
- Publication Date
- 2026-07-07
AI Technical Summary
In existing semantic communication systems, the end-to-end architecture leads to a mismatch between the semantic knowledge bases of the sending and receiving ends, resulting in performance degradation. This is especially true in communication scenarios where relay memory is limited or there are no relays. The knowledge base mismatch problem under non-nested updates is difficult to alleviate, and existing solutions require large amounts of data to be transmitted and have high latency, lacking flexibility.
The base station constructs a semantic adaptation layer, which is trained through an LSTM network. The training data is encoded using a transmitter model that matches the transmitter model and receiver model. The parameters are updated using the mean squared error as the loss function. The trained parameters are then sent to unmatched users to construct the semantic adaptation layer and achieve semantic communication.
It improves semantic communication performance in cases where the knowledge base is incomplete, reduces the number of transmission model parameters, avoids the impact of memory consumption, and improves communication performance and bandwidth utilization.
Smart Images
Figure CN121583243B_ABST
Abstract
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:
[0006] A semantic adaptation layer training method based on a transmitter model includes the following steps:
[0007] 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.
[0008] 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.
[0009] Step S3: After training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user. Each mismatched user constructs and enables its semantic adaptation layer based on the received parameters.
[0010] Furthermore, in step S1, all users communicating with the base station first complete the access to the base station and inform the base station of their receiving end model so that the base station can identify the mismatched users.
[0011] Furthermore, the semantic adaptation layer uses a cascaded network of LSTM network and fully connected layer.
[0012] Furthermore, each transmitting end model of the base station includes a DeepSC encoding network with a Transformer encoder, and each receiving end model of the user includes a DeepSC decoding network with a Transformer decoder.
[0013] Furthermore, in step S2, the channel constructed by the base station is an AWGN channel.
[0014] Furthermore, in step S2, the training process of the receiving end model and the sending end model includes nested updates and non-nested updates. Nested updates refer to training the initial model using training data during initial training, and training the previously trained model when the training data is updated. Non-nested updates refer to training the initial model even when the training data is updated.
[0015] Furthermore, in step S2, sentences with between 4 and 30 words from the English dataset of the European Parliament are used as the training data.
[0016] Furthermore, in step S2, the loss function is expressed as: ,in, For the second received signal, For the third received signal, For semantic adaptation layer parameters, To find the mean, The parameters of the sending model are: , For training data, This is channel noise.
[0017] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:
[0018] After identifying mismatched users, the base station of this invention constructs a semantic adaptation layer and trains the semantic adaptation layer for each mismatched user. During the training process, the base station encodes and transmits the training data using the transmitting end model used by the broadcast and the transmitting end model matched with the receiving end model of the mismatched user. The mean square error between the second and third received signals obtained from the two transmission lines is used as the training loss function. After training is completed, the base station packages 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 based on the received parameters. The addition of a semantic adaptation layer enables semantic communication even when the knowledge base is incomplete. During the training of the semantic adaptation layer, the receiving end model is not required. Therefore, the base station does not need to consume more memory due to the additional storage of receiving end models for each user, avoiding the adverse impact of memory consumption on communication performance. This effectively improves semantic communication performance. Furthermore, the semantic adaptation layer has a much smaller number of parameters compared to the DeepSC network. Therefore, when the semantic adaptation layer is trained, sending its parameters to unmatched users significantly reduces the number of model parameters required for transmission compared to sending the matched receiving end knowledge base to users, further improving communication performance. Attached Figure Description
[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] Figure 1 This is a flowchart of the present invention.
[0021] Figure 2 This is a schematic diagram of a base station broadcasting to multiple users.
[0022] Figure 3 This is a training logic block diagram of the present invention.
[0023] Figure 4 This invention provides a communication network architecture.
[0024] Figure 5 This is a communication system model based on the present invention.
[0025] Figure 6 This is a performance diagram of the present invention under nested updates.
[0026] Figure 7 This is a performance diagram of the present invention under non-nested update. Detailed Implementation
[0027] The present invention will be further described below through specific embodiments.
[0028] like Figure 1 As shown, the semantic adaptation layer training method based on the transmitter model includes the following steps:
[0029] Step S1: For mismatched users whose receiving model and the transmitting model used by the base station broadcast are different, the base station constructs a semantic adaptation layer and proceeds to step S2 to train the semantic adaptation layer. The semantic adaptation layer uses a cascaded network of 3 LSTM networks and 2 fully connected layers. The LSTM network has temporal memory capability and can learn the relationship between words in the sentence well, thereby effectively restoring the semantic communication performance.
[0030] Specifically, all users communicating with the base station first complete the access to the base station and register their own receiver model number (e.g., ...). Figure 1 Model A The system informs the base station, which then identifies the receiving end model for each user based on the number. It selects the transmitting end model corresponding to the receiving end model with the most users for broadcasting, and the remaining users are considered unmatched users. The base station only stores the receiving end model corresponding to each number.
[0031] In this invention, we define a semantic knowledge base as a semantic encoder and decoder model trained on a certain dataset through a certain number of training rounds, corresponding to the originating knowledge base (i.e., the originating model) and the receiving knowledge base (i.e., the receiving model), respectively. Semantic knowledge base matching means that the training process of the semantic encoder and decoder is completely consistent, that is, the dataset and training rounds are the same and the randomness during training is the same (e.g., a fixed random number seed); semantic knowledge base mismatch means that the training process of the semantic encoder and decoder is not completely consistent.
[0032] When a base station wants to communicate with a user via downlink semantic communication, it first needs to extract representational parameters based on the user's knowledge base and send them to the base station. At this point, the base station can identify a matching model from its stored model set based on these parameters, which is the identification of the semantic knowledge base.
[0033] A complete semantic knowledge base scenario is defined as one where the sending and receiving ends have matching models that can be used for communication. Otherwise, it is defined as an incomplete semantic knowledge base scenario. An incomplete semantic knowledge base will inevitably lead to semantic knowledge base mismatch. For example... Figure 2 As shown, the base station performs downlink broadcast semantic communication to two users, UE1 and UE2. The receiving end model of user UE1 is Model. A The receiving end model for UE2 users is Model B However, because base stations use a broadcast transmission method, they can only use one originating model. This will lead to a mismatch between the knowledge base and some users, for example, when using the originating model... ADuring transmission, the receiver model for UE2 users does not match. In this case, although the transmitter and receiver have matching models, they cannot use these models to achieve effective communication. Therefore, this scenario is also known as a semantic knowledge base incomplete scenario.
[0034] Each transmitting model of the base station includes a DeepSC encoding network with a Transformer encoder, and each receiving model of the user includes a DeepSC decoding network with a Transformer decoder. The DeepSC network is existing technology, proposed by Xie et al. in the article "Deeplearning enabled semantic communication systems" published in IEEE Transactions on Signal Processing, Volume 69, 2021.
[0035] 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.
[0036] Training logic diagram as follows Figure 3 As shown. During training, the channel constructed by the base station is an AWGN channel, and its channel noise is... The training data consisted of 73,472 sentences from the English dataset of the European Parliament, with 4 to 30 words per sentence. During training, an AWGN channel with a uniform SNR between 5 and 10 dB was used, and the number of training epochs was set to 80. The learning rate was... The batch size is 128, the optimizer used is the Adam optimizer, and its weight decay is set to... betas is set to (0.9, 0.98). The model's performance metric uses the Bilingual Evaluation Understudy (BLEU) score, with a default computational unit size of 1.
[0037] In training the originating and receiving models, the dataset is divided into three sets: the first 50,000 sentences are divided into datasets A, B, C, and a test set, containing 20,000, 10,000, 10,000, and 10,000 sentences respectively. The training process for the originating and receiving models can be either nested or non-nested. For nested updates, dataset A is used initially, and the model is trained based on a randomly initialized model. When the dataset is updated (i.e., dataset B is added), the model is retrained, but this time based on the previously trained model A, not the randomly initialized one. For non-nested updates, dataset A is used initially, and the model is trained based on a randomly initialized model. Even after dataset B is added, the model is still trained based on the previously initialized model. The specific processes for nested or non-nested updates are existing techniques.
[0038] In training the semantic adaptation layer, the training set is fixed at the first 40,000 sentences of the total dataset, i.e., the combined set of datasets A, B, and C. The input and output vector dimensions of the semantic adaptation layer are both 16. During training, the channel SNR is uniformly selected between 5 and 10 dB, the number of training epochs is set to 80, and the learning rate is [missing information]. The batch size is 256, the optimizer is Adam, the weight decay is set to 0, and the betas are set to (0.9, 0.999).
[0039] The parameters of the originating and receiving models are shown in Table 1, and the parameters of the semantic adaptation layer are shown in Table 2.
[0040] Table 1
[0041]
[0042] Table 2
[0043]
[0044] The loss function used for training is expressed as follows: ,in, For the second received signal, For the third received signal, For semantic adaptation layer parameters, To find the mean, The parameters of the sending model are: , For training data, For channel noise, For Model AHair, Model B The corresponding semantic adaptation layer is then implemented. During training, the parameters of both the sending and receiving models are fixed, while the parameters of the semantic adaptation layer are adjusted accordingly. Update.
[0045] Step S3: After training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user. Each mismatched user constructs and enables its semantic adaptation layer based on the received parameters.
[0046] The communication network constructed by combining the semantic adaptation layer is as follows Figure 4 As shown, the communication system model is as follows: Figure 5 As shown, the base station encodes and broadcasts communication data to each user. Matching users directly decode the data based on their receiving model, which matches the transmitting model used in the broadcast. Unmatching users process the received signal through a semantic adaptation layer before decoding it based on their receiving model. Figure 4 middle, For embedding vectors, E The dimension of the embedding vector corresponding to each word.
[0047] Figure 6 and Figure 7 In the comparison, Huffman+LDPC is used as the solution. For the DeepSC scheme, consider three model versions: versions 1, 2, and 3 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 using version 2 to communicate with version 1 in nested and non-nested update scenarios, respectively. "DeepSC(A+AB+ABC--A)—8 sym / word" and "DeepSC(ABC--A)—8 sym / word" represent the performance curves when using version 3 to communicate with version 1 in nested and non-nested update scenarios, respectively. "DeepSC(A+AB+ABC—A+AB)—8 sym / word" and "DeepSC(ABC--AB)—8 sym / word" represent the performance curves when using version 3 to communicate with version 2 in nested and non-nested update scenarios, respectively. The curves with the "with TMA" suffix represent the communication performance curves of the corresponding model pairs after adding the semantic adaptation layer obtained based on the training scheme proposed in this invention.
[0048] from Figure 6As can be seen, when using the nested update method, the present invention can enable the original mismatched model pair to achieve BLEU performance no less than that of the traditional scheme below 12.5dB, while having a bandwidth utilization gain of more than 1.5 times (in Baud / Hz) (that is, the number of symbols required to be transmitted can be compressed to less than 2 / 3 of the traditional scheme to achieve the same communication effect).
[0049] from Figure 7 It can be seen that when using the non-nested update method, the present invention can enable the original mismatched model pairs to achieve BLEU performance no less than that of the traditional scheme at less than 12dB, while having a bandwidth utilization of more than 1.5 times.
[0050] In this invention, the terms "first," "second," and "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. The use of terms such as "upper," "lower," "left," "right," "front," and "rear" to indicate orientation or positional relationships is based on the orientation or positional relationships shown in the accompanying drawings and is only for the convenience of describing the invention, not to indicate or imply that the device referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.
[0051] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0052] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.
Claims
1. A semantic adaptation layer training method based on a transmitter model, characterized in that: The steps include the following: 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 training is completed, the base station sends the parameters of each trained semantic adaptation layer to the corresponding mismatched user. Each mismatched user constructs and enables its semantic adaptation layer based on the received parameters.
2. The semantic adaptation layer training method based on the transmitter model according to claim 1, characterized in that: In step S1, all users communicating with the base station first complete the access to the base station and inform the base station of their receiving end model so that the base station can identify the mismatched users.
3. The semantic adaptation layer training method based on the transmitter model according to claim 1, characterized in that: The semantic adaptation layer uses a cascaded network of LSTM network and fully connected layer.
4. A semantic adaptation layer training method based on a transmitter model according to claim 1, 2, or 3, characterized in that: Each transmitting end model of the base station includes a DeepSC encoding network with a Transformer encoder, and each receiving end model of the user includes a DeepSC decoding network with a Transformer decoder.
5. A semantic adaptation layer training method based on a transmitter model according to claim 1, 2, or 3, characterized in that: In step S2, the channel constructed by the base station is an AWGN channel.
6. A semantic adaptation layer training method based on a transmitter model according to claim 1, 2, or 3, characterized in that: The training process of the receiving and sending models includes nested updates and non-nested updates. Nested updates refer to training the initial model using training data during initial training, and then training the previously trained model when the training data is updated. Non-nested updates refer to training the initial model even when the training data is updated.
7. A semantic adaptation layer training method based on a transmitter model according to claim 1, 2, or 3, characterized in that: In step S2, sentences with between 4 and 30 words from the English dataset of the European Parliament are used as the training data.
8. A semantic adaptation layer training method based on a transmitter model according to claim 1, 2, or 3, characterized in that: In step S2, the loss function is expressed as: ,in, For the second received signal, For the third received signal, For semantic adaptation layer parameters, To find the mean, The parameters of the sending model are: , For training data, This is channel noise.