CN-122001526-A - Structure-adaptive semantic communication method, device, equipment and storage medium

CN122001526ACN 122001526 ACN122001526 ACN 122001526ACN-122001526-A

Abstract

The invention discloses a semantic communication method, device and equipment with self-adaptive structure and a storage medium, belonging to the field of communication and artificial intelligence. The transmitting end combines the self and receiving end equipment capacity and task time delay and precision requirements to determine a target semantic model configuration strategy, activates a sub-network channel corresponding to the semantic coding and decoding model and exits branches in advance to form a target model, and transmits information after semantic coding and channel self-adaptive modulation. The receiving end determines a configuration strategy according to the indication information and the self capacity, configures a decoding model, demodulates and semantically decodes the recovered data. According to the invention, through dynamic adjustment of the model structure and joint encoding and decoding of the information source channels, the single model is adapted to terminals with different capacities, the problem of model adaptation of a fixed structure is solved, the dynamic balance of communication precision and time delay is achieved, and the robustness of the channels and the utilization efficiency of resources are improved.

Inventors

HAN SHUJUN
ZHU YUANJIA
ZHANG SHUNHANG
ZHANG WENZHAO
XU XIAODONG
SUN MENGYING
ZHANG PING

Assignees

北京邮电大学

Dates

Publication Date: 20260508
Application Date: 20251224

Claims (10)

1. The semantic communication method with self-adaptive structure is characterized by being applied to a transmitting end and comprising the following steps: Acquiring equipment capability information of a transmitting end and a receiving end and performance requirement information of a current communication task, wherein the equipment capability information comprises at least one of computing capability, storage capability and communication capability, and the performance requirement information comprises time delay constraint and precision constraint; determining a target semantic model configuration strategy based on the equipment capability information and the performance requirement information, wherein the target semantic model configuration strategy comprises width adjustable mechanism configuration information and depth adjustable mechanism configuration information of a semantic codec model; Activating or deactivating a sub-network channel with a corresponding width in the semantic coding and decoding model through a width adjusting factor according to the target semantic model configuration strategy, and enabling or disabling a corresponding early exit branch based on a depth adjusting threshold value to form a target semantic coding and decoding model; Carrying out semantic coding on the source data by utilizing the target semantic coding and decoding model to obtain semantic feature information; acquiring channel state information of a current channel, and modulating the semantic feature information through a signal-to-noise ratio self-adaptive modulation network according to the channel state information; and sending the modulated semantic feature information and the indication information of the target semantic model configuration strategy to a receiving end.
2. The architecture-adaptive semantic communication method according to claim 1, wherein the acquiring device capability information of a transmitting end and a receiving end specifically includes: the method comprises the steps of acquiring the processor utilization rate, the memory utilization ratio and the residual electric quantity parameters of a sending end in real time through a preset calculation power monitoring module, and calculating to obtain a first calculation power value of the sending end based on a setting algorithm; the receiving end returns a second calculation force value through a communication link, wherein the second calculation force value is obtained by the receiving end through the same calculation force monitoring module and calculation algorithm and is used for representing the equipment capacity level of the receiving end; The determining a target semantic model configuration strategy based on the equipment capability information and the performance requirement information specifically comprises the following steps: invoking a pre-stored strategy mapping table, wherein the strategy mapping table predefines the corresponding relation between different calculation force value intervals, different time delay constraint-precision constraint combinations, width adjustment factors and depth adjustment thresholds; and taking the smaller value of the first calculation force value of the transmitting end and the second calculation force value of the receiving end as a reference calculation force value, combining the time delay constraint and the precision constraint of the current communication task, and matching in a strategy mapping table to obtain a corresponding width adjustment factor and a depth adjustment threshold value so as to form a target semantic model configuration strategy.
3. The architecture-adaptive semantic communication method according to claim 1, wherein the forming a target semantic codec model according to a target semantic model configuration policy specifically comprises: Activating 25% of the number of channels of each network layer in the semantic coding and decoding model if the width adjustment factor in the target semantic model configuration strategy is 0.25, activating 50% of the number of channels if the width adjustment factor is 0.5, activating 75% of the number of channels if the width adjustment factor is 0.75, and activating 100% of the number of channels if the width adjustment factor is 1.0; Marking corresponding early exit branches in a middle layer of a semantic coding and decoding model according to depth adjustment thresholds in a target semantic model configuration strategy, wherein the early exit branches are in one-to-one correspondence with the depth adjustment thresholds, and the depth adjustment thresholds corresponding to the early exit branches are sequentially reduced from a network input layer to an output layer; the obtaining the channel state information of the current channel and modulating the semantic feature information specifically includes: transmitting a pilot signal to a receiving end through a channel estimation module, and calculating to obtain the signal-to-noise ratio of the current channel according to a signal attenuation value and an interference noise value fed back by the receiving end; Inputting the signal-to-noise ratio into an SNR adaptive modulation module of a signal-to-noise ratio adaptive modulation network, and outputting a high-dimensional modulation vector by the SNR adaptive modulation module; And carrying out dimension-by-dimension weighting on the semantic feature information by using the high-dimensional modulation vector to finish channel coding, and obtaining the modulated semantic feature information.
4. The semantic communication method with self-adaptive structure is characterized by being applied to a receiving end and comprising the following steps: receiving modulated semantic feature information sent by a sending end and indication information of a target semantic model configuration strategy; acquiring self equipment capability information of a receiving end, and determining a target semantic model configuration strategy of the receiving end by combining the indication information; Configuring a semantic coding and decoding model of a receiving end according to the target semantic model configuration strategy to form a target semantic decoding model; Transmitting a target semantic model configuration strategy of the receiving end back to the sending end for the sending end to optimize the target semantic coding model; obtaining channel state information of a current channel, and demodulating the received semantic feature information according to the channel state information; And carrying out semantic decoding on the demodulated semantic feature information by using the target semantic decoding model, and recovering the original communication data.
5. The architecture-adaptive semantic communication method according to claim 4, wherein the determining, in combination with the indication information, a target semantic model configuration policy of a receiving end specifically includes: Analyzing a first width adjustment factor and a first depth adjustment threshold determined by a transmitting end from the indication information; Based on a second calculation force value of the receiving end, a second width adjustment factor and a second depth adjustment threshold corresponding to the receiving end are obtained through policy mapping table matching; And taking the smaller value of the first width adjustment factor and the second width adjustment factor as a target width adjustment factor of the receiving end, and taking the smaller value of the first depth adjustment threshold and the second depth adjustment threshold as a target depth adjustment threshold of the receiving end to form a target semantic model configuration strategy of the receiving end.
6. The architecture-adaptive semantic communication method according to claim 4, wherein the performing semantic decoding on the demodulated semantic feature information by using the target semantic decoding model specifically comprises: inputting the demodulated semantic feature information into a target semantic decoding model, and sequentially carrying out feature conversion through each intermediate layer; Predicting the output characteristics of the current layer at the position of the early exit branch corresponding to each middle layer through a classifier built in the branch to obtain a prediction result, and calculating the information entropy of the prediction result; If the information entropy is larger than or equal to the target depth adjustment threshold value of the receiving end, the feature conversion of the subsequent network layer is stopped, the current prediction result is output, and if the information entropy is larger than or equal to the target depth adjustment threshold value of the receiving end, the feature is continuously input into the next middle layer until the network output layer or the early exit condition is met.
7. A structure-adaptive semantic communication device, which adopts the structure-adaptive semantic communication method according to any one of claims 1 to 3, and is characterized in that the structure-adaptive semantic communication device is applied to a transmitting end, and comprises: The device capability and demand acquisition module is used for acquiring device capability information of a sending end and a receiving end and performance demand information of a current communication task, wherein the device capability information comprises at least one of computing capability, storage capability and communication capability, and the performance demand information comprises time delay constraint and precision constraint; The first strategy determining module is used for determining a target semantic model configuration strategy based on the equipment capability information and the performance requirement information, wherein the target semantic model configuration strategy comprises width adjustable mechanism configuration information and depth adjustable mechanism configuration information of a semantic codec model; The first model configuration module is used for activating or deactivating a sub-network channel with a corresponding width in the semantic codec model through a width adjustment factor according to the target semantic model configuration strategy, and enabling or disabling a corresponding early exit branch based on a depth adjustment threshold value to form a target semantic codec model; the semantic coding module is used for carrying out semantic coding on the source data by utilizing the target semantic coding and decoding model to obtain semantic characteristic information; The first semantic feature modulation module is used for acquiring channel state information of a current channel, and modulating the semantic feature information through a signal-to-noise ratio adaptive modulation network according to the channel state information; the indication information sending module is used for sending the modulated semantic feature information and the indication information of the target semantic model configuration strategy to the receiving end.
8. A structure-adaptive semantic communication apparatus, which adopts the structure-adaptive semantic communication method according to any one of claims 4 to 6, and is characterized by being applied to a receiving end, comprising: The indication information receiving module is used for receiving the modulated semantic feature information sent by the sending end and the indication information of the target semantic model configuration strategy; the second strategy determining module is used for acquiring the equipment capacity information of the receiving end and determining a target semantic model configuration strategy of the receiving end by combining the indication information; The second model configuration module is used for configuring a semantic coding and decoding model of the receiving end according to the target semantic model configuration strategy to form a target semantic decoding model; The configuration strategy feedback module is used for transmitting the target semantic model configuration strategy of the receiving end back to the sending end and optimizing the target semantic coding model by the sending end; the second semantic feature modulation module is used for acquiring channel state information of a current channel and demodulating the received semantic feature information according to the channel state information; And the semantic decoding module is used for carrying out semantic decoding on the demodulated semantic feature information by utilizing the target semantic decoding model and recovering the original communication data.
9. A structurally adaptive semantic communication device comprising a processor, a memory and a program or instruction stored on the memory and executable on the processor, the processor implementing the structurally adaptive semantic communication method of any one of claims 1-6 when executing the program or instruction.
10. A computer readable storage medium, characterized in that the storage medium has stored thereon a program or instructions which, when executed by a processor, implement the steps of the architecture-adaptive semantic communication method of any of claims 1-6.

Description

Structure-adaptive semantic communication method, device, equipment and storage medium Technical Field The invention belongs to the technical field of communication technology and artificial intelligence, and particularly relates to a semantic communication method, device and equipment with self-adaptive structure and a storage medium. Background Semantic communication is a new paradigm of next-generation communication, and has advantages of high transmission reliability and high transmission efficiency. Wherein the semantic codec is an important component in a semantic communication system. However, the existing semantic codec structure often has a fixed network structure, and it is difficult to meet different demands of different users on communication task time delay and precision. Meanwhile, for information carriers such as images and videos with rich semantic features, the requirement on computing capacity and storage space is greatly improved due to the fact that a semantic codec with a more complex structure is needed, and timeliness of communication tasks is difficult to guarantee. Therefore, the existing semantic communication codec model of the fixed architecture is difficult to meet different communication requirements of all users. Disclosure of Invention In view of the above, the present invention is directed to a semantic communication method, device, apparatus and storage medium with adaptive structure, so as to solve or partially solve the problems mentioned in the background art. Based on the above object, in a first aspect, the present invention provides a semantic communication method with adaptive structure, which is applied to a transmitting end, and includes: Acquiring equipment capability information of a transmitting end and a receiving end and performance requirement information of a current communication task, wherein the equipment capability information comprises at least one of computing capability, storage capability and communication capability, and the performance requirement information comprises time delay constraint and precision constraint; determining a target semantic model configuration strategy based on the equipment capability information and the performance requirement information, wherein the target semantic model configuration strategy comprises width adjustable mechanism configuration information and depth adjustable mechanism configuration information of a semantic codec model; Activating or deactivating a sub-network channel with a corresponding width in the semantic coding and decoding model through a width adjusting factor according to the target semantic model configuration strategy, and enabling or disabling a corresponding early exit branch based on a depth adjusting threshold value to form a target semantic coding and decoding model; Carrying out semantic coding on the source data by utilizing the target semantic coding and decoding model to obtain semantic feature information; acquiring channel state information of a current channel, and modulating the semantic feature information through a signal-to-noise ratio self-adaptive modulation network according to the channel state information; and sending the modulated semantic feature information and the indication information of the target semantic model configuration strategy to a receiving end. As a preferred scheme of the semantic communication method with adaptive structure, the acquiring the device capability information of the sending end and the receiving end specifically includes: the method comprises the steps of acquiring the processor utilization rate, the memory utilization ratio and the residual electric quantity parameters of a sending end in real time through a preset calculation power monitoring module, and calculating to obtain a first calculation power value of the sending end based on a setting algorithm; And the receiving end returns a second calculation force value through a communication link, wherein the second calculation force value is obtained by the receiving end through the same calculation force monitoring module and calculation algorithm and is used for representing the equipment capacity level of the receiving end. As a preferred scheme of the structure-adaptive semantic communication method, the determining a target semantic model configuration policy based on the device capability information and the performance requirement information specifically includes: invoking a pre-stored strategy mapping table, wherein the strategy mapping table predefines the corresponding relation between different calculation force value intervals, different time delay constraint-precision constraint combinations, width adjustment factors and depth adjustment thresholds; and taking the smaller value of the first calculation force value of the transmitting end and the second calculation force value of the receiving end as a reference calculation force value, combining the time delay constraint and t