A multi-user semantic knowledge base synchronization method based on federated distillation learning
Patent Information
- Application Number
- CN202410085493.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-22
- Publication Date
- 2026-10-09
- Estimated Expiration
- 2044-01-22
AI Technical Summary
[0003]目前，大量研究仅讨论了端到端语义通信中的知识库同步问题，未考虑语义通信网络的场景下多用户的知识库同步问题
[0048] (1) To address the communication error problem caused by the mismatch of knowledge bases among multiple users in semantic communication, semantic knowledge base synchronization is achieved by cooperating with the server at the device end where the knowledge bases are mismatched. Semantic knowledge base synchronization among multiple users is realized, and semantic encoding and decoding are completed based on the synchronized knowledge base, thereby improving the accuracy of semantic communication transmission.
Smart Images
Figure CN117993491B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of semantic communication, specifically relating to a multi-user semantic knowledge base synchronization method based on federated distillation learning. Background Technology
[0002] In recent years, with the rapid development of artificial intelligence and deep learning technologies, research on semantic communication based on artificial intelligence technology has been vigorously carried out to further improve communication efficiency. Semantic communication utilizes neural network models to extract and transmit the meaning contained in data, removing redundant information, effectively improving communication efficiency, and exhibiting better resistance to channel noise. Among these components, the semantic knowledge base, as a crucial element of the semantic communication system, guides the transmitting and receiving ends in semantic encoding and decoding of transmitted information. However, when the background knowledge bases of communication devices are inconsistent, the receiving end will be unable to correctly perform semantic decoding of the sent information, leading to misunderstandings and severely degrading the performance of the communication system. Therefore, it is urgent to solve the problem of semantic knowledge base synchronization.
[0003] Currently, much research has only discussed knowledge base synchronization in end-to-end semantic communication, neglecting the multi-user knowledge base synchronization problem in semantic communication networks. However, multi-user semantic communication is far larger in scale than end-to-end communication systems, requiring significantly more computational and communication resources for knowledge base synchronization. Therefore, designing efficient knowledge base synchronization methods for semantic communication networks presents a new challenge for the development of semantic communication. Summary of the Invention
[0004] To address the shortcomings and improvement needs of existing technologies, the purpose of this invention is to provide a multi-user semantic knowledge base synchronization method based on federated distillation learning, which realizes the synchronization of multi-user semantic communication knowledge bases, saves the communication resources required for knowledge base synchronization, and improves the accuracy of semantic communication.
[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution.
[0006] 1. A method for synchronizing a multi-user semantic knowledge base based on federated distillation learning, comprising the following steps:
[0007] S1. The server initializes the structure of the global knowledge base model, which includes a semantic encoder network and a semantic decoder network. The parameters are initialized in a random manner.
[0008] S2. Let D be the set of all devices. The server randomly selects a portion of the devices to form a set D' for the next global knowledge base model aggregation. At the same time, the server distributes the existing global knowledge base model to the devices in the selected set.
[0009] S3. The device receives the global knowledge base model from the server and updates the local semantic knowledge base, including the local teacher model and student model.
[0010] S4. Each device in the set of devices selected by the server transmits the updated student knowledge base model to the server.
[0011] S5. After receiving the student knowledge base model sent by all devices in the set, the server performs a global knowledge base aggregation update and tests and verifies the synchronization degree of the knowledge base.
[0012] S6. Repeat step S2 until the knowledge base synchronization level reaches the set threshold or the number of communication rounds between the server and the client reaches the set threshold T, then stop updating and proceed to step S7.
[0013] S7. Distribute the global knowledge base model, which has been synchronized with the knowledge base on the server side, to all clients.
[0014] Furthermore, the structure of the global knowledge base model consists of the encoder-decoder structure of the Transformer model, and the parameters are initialized in a random manner.
[0015] The semantic encoder takes a piece of English text as input, extracts semantic information from the English text under the guidance of a global knowledge base, and encodes the English text into a feature vector through a neural network.
[0016] The semantic decoder takes a feature vector as input and, guided by a global knowledge base, converts the feature vector into English text.
[0017] Furthermore, for step S3, each device has a private local knowledge base model as the teacher model and a shared global knowledge base as the student model.
[0018] The teacher model and the student model learn from each other through knowledge distillation. The teacher model can extract more knowledge from the training data and guide the training of the student model.
[0019] By learning knowledge from the teacher model and local data, the student model can achieve higher semantic encoding and decoding accuracy in a limited number of training iterations;
[0020] Due to differences in computing and storage resources, the size and structure of the teacher semantic knowledge base model vary among devices, but the global knowledge base structure of each device must remain consistent.
[0021] Furthermore, step S3 specifically includes:
[0022] In the t-th communication round, the devices in set D' receive the global knowledge base model from the server. The device first uses the global knowledge base model Assigned to the local knowledge base student model Right now
[0023]
[0024] Then, the local semantic knowledge base is updated, and the specific process is as follows:
[0025] Device i trains a semantic knowledge base based on the local dataset. Under the guidance of the semantic knowledge base, the local data will extract semantic information, realize the semantic encoding of the data, and generate a semantic representation vector.
[0026] The device's semantic knowledge base includes a large teacher model and a small student model. The large teacher model and the small student model learn from each other through knowledge distillation. The large teacher model can extract more knowledge from the training data and guide the training of the small student model. The small student model can achieve higher semantic encoding and decoding accuracy in a limited number of training iterations by learning knowledge from the teacher model and local data.
[0027] In addition, each device may have different computing and storage resources, and the resources are heterogeneous. The local teacher model structure can be set according to the local resources, and the size of the teacher semantic knowledge base model can be different for different devices.
[0028] Device i collaboratively trains the knowledge base teacher model based on local data. With knowledge base student model The local data of the training set is used as input, and semantic encoding is performed by teacher and student models to obtain semantic feature vectors. Then, channel coding is performed to further compress the data. The data is then transmitted through the channel through numerical simulation. The data output by the channel is then channel decoded, and the channel decoding result is then semantically decoded to obtain the data received by the receiving end.
[0029] Furthermore, step S3 also includes the following calculations:
[0030] The neural networks of the teacher and student models in the semantic knowledge base employ forward propagation. Then, the loss functions of the teacher and student models are calculated separately, followed by backpropagation to update the model parameters. Iterative training improves the accuracy of the semantic knowledge base model on device i. Specifically, the device calculates the loss function L of the teacher model based on cross-entropy loss and distillation loss in the following way. t,i :
[0031]
[0032]
[0033]
[0034] Where CE() represents the cross-entropy loss, KL() represents the KL divergence, i.e., the distillation loss, and y i The label representing the true result, y t,i y represents the predicted label of the teacher's network output. s,i This represents the predicted labels for the student's network output;
[0035] The device calculates the loss function of the student model based on cross-entropy loss and distillation loss in the following way:
[0036]
[0037]
[0038]
[0039] Where CE() represents the cross-entropy loss, KL() represents the KL divergence, i.e., the distillation loss, and y i The label representing the true result, y t,i y represents the predicted label of the teacher's network output. s,i This represents the predicted label of the student's network output.
[0040] Furthermore, in step S5, after the server receives the student knowledge base models transmitted through the channel from all devices in the set, it performs a global knowledge base aggregation update, calculating a weighted average of the network parameters of the knowledge base models to obtain the updated global knowledge base. It can be represented as:
[0041]
[0042] Furthermore, in step S6, the knowledge base synchronization level is calculated by the global knowledge base model. When the transmitted data is text, the semantic similarity between two sentences is measured using sentence similarity based on the BERT model. The semantic similarity is defined as follows:
[0043]
[0044] Where s is the sentence sent. B represents the sentence received after semantic communication transmission guided by a semantic knowledge base, and B() is a pre-trained BERT model for extracting semantic information from the text. The semantic similarity value ranges from [0,1], with higher similarity indicating better synchronization of the knowledge base.
[0045] Furthermore, once the global knowledge base aggregation update on the server side is completed, the synchronization level of the knowledge base is immediately calculated. If the set threshold is reached, or the number of communication rounds between the server and the client reaches the set threshold T, the update is stopped, meaning that the global knowledge base model has been trained and updated.
[0046] After synchronization is complete, the server distributes the synchronized global semantic knowledge base to all devices as follows: the server and client continuously communicate to update the global knowledge base. When the synchronization level of the global knowledge base model reaches the requirement or the number of communication rounds between the server and the client reaches the set threshold, the global knowledge base synchronization is considered complete. The server distributes the synchronized global knowledge base to all clients, and the clients conduct semantic communication through the global knowledge base.
[0047] Beneficial effects: Compared with the prior art, the substantial progress and significant effects of the present invention include:
[0048] (1) To address the communication error problem caused by the mismatch of knowledge bases among multiple users in semantic communication, semantic knowledge base synchronization is achieved by cooperating with the server at the device end where the knowledge bases are mismatched. Semantic knowledge base synchronization among multiple users is realized, and semantic encoding and decoding are completed based on the synchronized knowledge base, thereby improving the accuracy of semantic communication transmission.
[0049] (2) To address the problem of high communication overhead between the client and the server, this invention designs a framework for knowledge distillation on the client side. Each client has a large teacher model and a small student model locally. The two models are updated collaboratively. Through the knowledge distillation method, after a certain number of communication rounds, the student model can achieve a knowledge base synchronization accuracy similar to that of the teacher model. Moreover, when the device communicates with the server, only the small student model is transmitted, which reduces the communication overhead required for synchronization.
[0050] (3) To address the issue of heterogeneous client resources, this invention introduces a knowledge distillation mechanism and designs a framework for collaborative updating of teacher and student models. Different large-scale teacher network structures are designed for clients with heterogeneous resources to adapt to the resource heterogeneity of the devices. Simultaneously, each client's student network has the same dimensions and structure. When the server performs model aggregation, it averages the student network parameters, which solves the problem of the inability to aggregate knowledge base models with different structures and dimensions. Attached Figure Description
[0051] Figure 1 A flowchart of a multi-user semantic knowledge base synchronization method based on federated distillation learning;
[0052] Figure 2 This diagram illustrates local updates performed on the device via knowledge distillation. Detailed Implementation
[0053] The present invention will now be described in detail with reference to the accompanying drawings and specific examples.
[0054] This invention proposes a multi-user semantic knowledge base synchronization method based on federated distillation learning. This method includes multiple devices synchronizing their semantic knowledge bases and a server. Each device possesses local private data and trains its local semantic knowledge base based on this data. However, the data differs between devices, resulting in inconsistent local semantic knowledge bases and hindering semantic communication. Semantic knowledge base synchronization is necessary for efficient semantic communication. Devices can be computers or other devices with sufficient computing, storage, and communication resources. The server globally aggregates the local semantic knowledge bases from multiple devices, constructs and updates a global semantic knowledge base, and guides each device to update and synchronize its knowledge base. The server must possess sufficient computing, storage, and communication resources.
[0055] Based on the above technical solutions, this method is mainly implemented in the following manner:
[0056] The server performs global semantic knowledge base initialization.
[0057] The server sends the knowledge base to be synchronized to the device. After receiving the data, the device performs a local update. The device then sends the updated global knowledge base back to the server.
[0058] The server performs global semantic knowledge base aggregation.
[0059] After synchronization is complete, the server distributes the synchronized global semantic knowledge base to all devices.
[0060] Furthermore, the implementation of this method includes the following steps:
[0061] (1) Parameter initialization.
[0062] The server initializes the semantic global knowledge base model structure and performs parameter initialization. The global knowledge base model includes a semantic encoder network and a semantic decoder network, and the parameters are initialized randomly.
[0063] (2) The server distributes the knowledge base to be synchronized.
[0064] The server randomly selects a subset of devices to participate in the next global knowledge base model aggregation and distributes the existing global semantic knowledge base model to the selected set of devices.
[0065] (3) Local update on the device.
[0066] Each device has a private local knowledge base model as the teacher model and a shared global knowledge base as the student model. The teacher and student models learn from each other through knowledge distillation. The teacher model can extract more knowledge from the training data and guide the student model's training; the student model, by learning from the teacher model and local data, can achieve higher semantic encoding / decoding accuracy within a limited number of training iterations. While the size and structure of the teacher semantic knowledge base model may differ across devices due to variations in computing and storage resources, the global knowledge base structure must remain consistent across all devices.
[0067] (4) Upload the knowledge base on the device.
[0068] After all devices in the collection complete the local update process, they upload the global knowledge base model to the server.
[0069] (5) Server aggregated knowledge base.
[0070] After receiving the student knowledge base models transmitted through the channel from all devices in the set, the server performs a global knowledge base aggregation update by calculating a weighted average of the network parameters of the knowledge base models, thereby obtaining the updated global knowledge base.
[0071] (6) The server distributes the synchronized global knowledge base.
[0072] Repeat steps (2)-(5) above until the synchronization level of the global knowledge base model reaches the required level or the number of communication rounds between the server and the client reaches the set threshold. The global knowledge base synchronization is then considered complete. The server distributes the synchronized global knowledge base to all clients, and the clients use the global knowledge base for semantic communication.
[0073] Combination Figure 1 As shown, the present invention provides a flowchart of a multi-user semantic knowledge base synchronization method based on federated distillation learning. The processing flow of this method includes the following steps:
[0074] S1. The server initializes the global knowledge base model structure and performs parameter initialization.
[0075] In one feasible implementation, the global knowledge base model initialized on the server side includes a semantic encoder and a semantic decoder. Specifically, the input to the semantic encoder is a piece of English text, which may contain one or more sentences, each containing multiple English words. Guided by the global knowledge base, the semantic encoder extracts the meaning, i.e., semantic information, from the English text and encodes it into a feature vector using a neural network. The input to the semantic decoder is a feature vector containing rich semantic information. Guided by the global knowledge base, the semantic decoder reconstructs the English text from this feature vector. The structure of the global knowledge base can consist of the encoder-decoder structure of a Transformer model, with parameters initialized randomly.
[0076] S2. The server randomly selects a subset of devices to form set D' for the next global knowledge base model aggregation. Simultaneously, the server distributes the existing global knowledge base model to the devices in the selected set.
[0077] Let D be the set of all devices. The server randomly selects a subset of devices to form this set. Furthermore, all devices in D' will participate in the next aggregation of the global knowledge base model. The server will then distribute the existing global knowledge base model to all devices in the set.
[0078] S3. The device receives the global knowledge base model from the server and updates the local semantic knowledge base, including the local teacher model and student model.
[0079] In the t-th communication round, the devices in set D' receive the global knowledge base model from the server. The device first uses the global knowledge base model Assigned to the local knowledge base student model Right now
[0080]
[0081] Then, the local semantic knowledge base is updated, and the specific process is as follows:
[0082] Device i trains a semantic knowledge base based on a local dataset. Semantic information is extracted from the local data under the guidance of this knowledge base. Semantic encoding of the data is then implemented, generating semantic representation vectors. The device's semantic knowledge base includes a large teacher model and small student models. These models learn from each other through knowledge distillation. The large teacher model extracts more knowledge from the training data and guides the training of the small student models. The small student models, by learning from the teacher model and the local data, achieve higher semantic encoding / decoding accuracy within a limited number of training iterations.
[0083] Each device can have different computing and storage resources, and these resources are heterogeneous. The local teacher model structure can be set according to local resources, and the size of the teacher semantic knowledge base model can be different for different devices.
[0084] Device i collaboratively trains the knowledge base teacher model based on local data. With knowledge base student model The local data of the training set is used as input, and semantic encoding is performed by teacher and student models to obtain semantic feature vectors. Then, channel coding is performed to further compress the data. The data is then transmitted through the channel through numerical simulation. The data output by the channel is then channel decoded, and the channel decoding result is then semantically decoded to obtain the data received by the receiving end.
[0085] The neural networks of the teacher and student models in the semantic knowledge base employ forward propagation. Then, the loss functions of the teacher and student models are calculated separately, followed by backpropagation to update the model parameters. Iterative training improves the accuracy of the semantic knowledge base model on device i. Specifically, the device calculates the loss function L of the teacher model based on cross-entropy loss and distillation loss in the following manner. t,i :
[0086]
[0087]
[0088]
[0089] Where CE() represents the cross-entropy loss, KL() represents the KL divergence, i.e., the distillation loss, and y i The label representing the true result, y t,i y represents the predicted label of the teacher's network output. s,i This represents the predicted label output by the student network. The device calculates the loss function of the student model based on cross-entropy loss and distillation loss in the following way:
[0090]
[0091]
[0092]
[0093] Where CE() represents the cross-entropy loss, KL() represents the KL divergence, i.e., the distillation loss, and y i The label representing the true result, y t,i y represents the predicted label of the teacher's network output. s,i This represents the predicted label of the student's network output.
[0094] S4. Each device in the set of devices selected by the server transmits the updated student knowledge base model to the server.
[0095] After device i in set D' completes local training, the student model will be... Upload to the server to update the global knowledge base model.
[0096] S5. After receiving the student knowledge base models sent by all devices in the set, the server performs a global knowledge base aggregation update and tests and verifies the synchronization level of the knowledge base.
[0097] After receiving the student knowledge base models transmitted through the channel from all devices in the set, the server performs a global knowledge base aggregation update. This involves calculating a weighted average of the network parameters of the knowledge base models to obtain the updated global knowledge base. It can be represented as:
[0098]
[0099] S6. Repeat S2 until the knowledge base synchronization level reaches the set threshold or the number of communication rounds between the server and the client reaches the set threshold T, then stop updating and execute S7.
[0100] The knowledge base synchronization level is calculated by the global knowledge base model. Knowledge base synchronization level, or knowledge base synchronization accuracy, refers to evaluating the semantic similarity between the input data and the data obtained after semantic communication transmission. When the transmitted data is text, sentence similarity based on the BERT model is used to measure the semantic similarity between two sentences. Semantic similarity is defined as follows:
[0101]
[0102] Where s is the sentence sent. B represents the sentence received after semantic communication transmission guided by a semantic knowledge base, and B() is a pre-trained BERT model for extracting semantic information from the text. The semantic similarity value ranges from [0,1], with higher similarity indicating better synchronization of the knowledge base.
[0103] Once the global knowledge base aggregation update is completed on the server side, the current knowledge base synchronization level is immediately calculated. If the set threshold is reached, or the number of communication rounds between the server and the client reaches the set threshold T, then the update is stopped, meaning that the global knowledge base model has been trained and updated.
[0104] S7: Distribute the global knowledge base model, which has been synchronized on the server side, to all clients.
[0105] The server distributes a synchronized global semantic knowledge base to all devices in set D. Once the client receives the global semantic knowledge base, it can use it for semantic communication.
Claims
1. A method for synchronizing a multi-user semantic knowledge base based on federated distillation learning, characterized by the following steps: include: S1. The server initializes the structure of the global knowledge base model, which includes a semantic encoder network and a semantic decoder network. The parameters are initialized in a random manner. S2. Let D be the set of all devices. The server randomly selects a portion of the devices to form the set. Perform the next global knowledge base model aggregation; at the same time, the server distributes the existing global knowledge base models to the devices in the selected set; S3. The device receives the global knowledge base model from the server and updates the local semantic knowledge base, including the local teacher model and student model. Step S3 specifically includes: In the t-th communication round, the set The device receives the global knowledge base model from the server. The device first uses the global knowledge base model Assigned to the local knowledge base student model ,Right now Then, the local semantic knowledge base is updated, and the specific process is as follows: Device i trains a semantic knowledge base based on the local dataset. Under the guidance of the semantic knowledge base, the local data will extract semantic information, realize the semantic encoding of the data, and generate a semantic representation vector. The device's semantic knowledge base includes a large teacher model and a small student model. The large teacher model and the small student model learn from each other through knowledge distillation. The large teacher model can extract more knowledge from the training data and guide the training of the small student model. The small student model can achieve higher semantic encoding and decoding accuracy in a limited number of training iterations by learning knowledge from the teacher model and local data. In addition, each device has different computing and storage resources, and the resources are heterogeneous. The local teacher model structure is set according to the local resources, and the size of the teacher semantic knowledge base model is different for different devices. Device i collaboratively trains the knowledge base teacher model based on local data. With knowledge base student model The local data of the training set is used as input, and semantic encoding is performed by the teacher and student models to obtain semantic feature vectors. Then, channel coding is performed to further compress the data. Then, the data is transmitted through the channel through numerical simulation. The data output by the channel is channel decoded, and the channel decoding result is semantically decoded to obtain the data received by the receiving end. S4. Each device in the set of devices selected by the server transmits the updated student knowledge base model to the server. S5. After receiving the student knowledge base model sent by all devices in the set, the server performs a global knowledge base aggregation update and tests and verifies the synchronization degree of the knowledge base. S6. Repeat step S2 until the knowledge base synchronization level reaches the set threshold or the number of communication rounds between the server and the client reaches the set threshold T, then stop updating and proceed to step S7. S7. Distribute the global knowledge base model, which has been synchronized with the knowledge base on the server side, to all clients.
2. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, The structure of the global knowledge base model consists of the encoder-decoder structure of the Transformer model, and the parameters are initialized in a random manner.
3. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, The semantic encoder takes a piece of English text as input, extracts semantic information from the English text under the guidance of a global knowledge base, and encodes the English text into a feature vector through a neural network. The semantic decoder takes a feature vector as input and, guided by a global knowledge base, converts the feature vector into English text.
4. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, For step S3, each device has a private local knowledge base model as the teacher model and a shared global knowledge base as the student model. The teacher model and the student model learn from each other through knowledge distillation. The teacher model can extract more knowledge from the training data and guide the training of the student model. By learning knowledge from the teacher model and local data, the student model can achieve higher semantic encoding and decoding accuracy in a limited number of training iterations; Due to differences in computing and storage resources, the size and structure of the teacher semantic knowledge base model vary among devices, but the global knowledge base structure of each device must remain consistent.
5. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, Step S3 also includes the following calculations: The neural networks of the teacher and student models in the semantic knowledge base employ forward propagation. Then, the loss functions of the teacher and student models are calculated separately, followed by backpropagation to update the model parameters. Iterative training improves the accuracy of the semantic knowledge base model on device i. Specifically, the device calculates the loss function of the teacher model based on cross-entropy loss and distillation loss in the following way. : in, CE( ) Represents cross-entropy loss, KL( ) This represents the KL divergence, i.e., distillation loss. Labels representing the true results This represents the predicted labels output by the teacher's network. Represents the predicted labels for the student's network output; The device calculates the loss function of the student model based on cross-entropy loss and distillation loss in the following way: in, CE( ) Represents cross-entropy loss, KL( ) This represents the KL divergence, i.e., distillation loss. Labels representing the true results This represents the predicted labels output by the teacher's network. This represents the predicted label of the student's network output.
6. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, In step S5, after receiving the student knowledge base models transmitted through the channel from all devices in the set, the server performs a global knowledge base aggregation update, calculating a weighted average of the network parameters of the knowledge base models to obtain the updated global knowledge base. , represented as: 。 7. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, In step S6, the knowledge base synchronization level is calculated by the global knowledge base model. When the transmitted data is text, the semantic similarity between two sentences is measured using sentence similarity based on the BERT model. The semantic similarity is defined as follows: in, For the sentence sent, The sentence received after semantic communication transmission guided by a semantic knowledge base. B( ) This is a pre-trained BERT model for extracting semantic information from text; the semantic similarity value is [0,1], and the higher the similarity, the better the synchronization of the knowledge base.
8. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 7, characterized in that, Once the global knowledge base aggregation update on the server side is completed, the current knowledge base synchronization level is immediately calculated. If the set threshold is reached, or the number of communication rounds between the server and the client reaches the set threshold T, the update is stopped, meaning that the global knowledge base model has been trained and updated.
9. The multi-user semantic knowledge base synchronization method based on federated distillation learning according to claim 1, characterized in that, After synchronization is complete, the server distributes the synchronized global semantic knowledge base to all devices as follows: the server and client continuously communicate to update the global knowledge base. When the synchronization level of the global knowledge base model reaches the requirement or the number of communication rounds between the server and the client reaches the set threshold, the global knowledge base synchronization is considered complete. The server distributes the synchronized global knowledge base to all clients, and the clients conduct semantic communication through the global knowledge base.