Search

CN-121619601-B - Semantic communication method based on deep learning and energy consumption optimization

CN121619601BCN 121619601 BCN121619601 BCN 121619601BCN-121619601-B

Abstract

The invention discloses a semantic communication method based on deep learning and energy consumption optimization, which comprises the steps of S1, constructing a deep neural network semantic communication system based on a super prior structure, S2, carrying out energy consumption modeling on the deep neural network semantic communication system, wherein the energy consumption modeling comprises information source coding and decoding energy consumption modeling and communication transmission energy consumption modeling, and S3, constructing a joint optimization problem by taking model reasoning complexity, calculation frequency and wireless transmission power as adjustable decision variables, and solving to obtain a communication strategy. The invention constructs a unified calculation energy consumption and communication energy consumption model, and on the premise of meeting the end-to-end time delay constraint, the invention jointly optimizes key parameters such as complexity of semantic source coding model, operating frequency of processor, wireless transmission power and the like, thereby realizing the minimization of the total energy consumption of the system.

Inventors

  • HUANG CHUAN
  • YUAN KAI

Assignees

  • 电子科技大学(深圳)高等研究院

Dates

Publication Date
20260508
Application Date
20260130

Claims (7)

  1. 1. A semantic communication method based on deep learning and energy consumption optimization is characterized by comprising the following steps: Step S1, constructing a deep neural network semantic communication system based on a super prior structure, wherein the deep neural network semantic communication system comprises a transmitting end and a receiving end; The transmitting end comprises a depth source coder and a channel coder, wherein in the depth source coder, a deep neural network is utilized to extract and compress semantic features of source data, and after digital channel coding of the channel coder, a wireless channel is utilized to transmit; the receiving end comprises a depth information source decoder and a channel decoder, and corresponding information source data is recovered through the depth information source decoder after the channel decoder is utilized to carry out channel decoding on the received signals; S2, carrying out energy consumption modeling on a deep neural network semantic communication system, wherein the energy consumption modeling comprises information source coding and decoding energy consumption modeling and communication transmission energy consumption modeling; S3, constructing a joint optimization problem by taking model reasoning complexity, calculation frequency and wireless transmission power as adjustable decision variables, and solving to obtain a communication strategy; The step S3 includes: s301, taking model reasoning complexity, calculation frequency and wireless transmission power as adjustable decision variables, and constructing a joint optimization problem: Total energy consumption With end-to-end delay Expressed as: ; ; Given maximum allowable end-to-end delay The optimization objective is to minimize the total energy consumption while satisfying the delay constraint, and the optimization problem is expressed as: ; Wherein, the To calculate the energy consumption; Searching decision variable for average transmission delay and optimizing problem 、 And Is configured to have an objective function Taking the minimum, the constraint indicates that the end-to-end delay does not exceed a given threshold, and in the optimization problem, Controlling the number of source encoded output bits And calculating the time delay ; Determining reasoning calculation time delay and energy consumption; At the same time influence shannon rate And communication energy consumption In a constraint of Representing user experience or system real-time requirements; s302, fixing part of variables in the model, gradually analyzing the coupling relation between decision variables, and deriving a closed solution by using a scaling law and an energy efficiency model: (1) Fixed model complexity And calculating the frequency When the problem is degenerated to the transmit power Single variable optimization of (c) with the aim of minimizing communication energy consumption Equivalent to maximizing energy efficiency given delay constraints Wherein the energy efficiency is The definition is as follows: ; Wherein the method comprises the steps of Taking the modulus of h and energy efficiency I.e. the number of significant bits that can be transmitted per unit energy, by combining For the following Obtaining the optimal transmitting power by deriving and enabling the derivative to be zero; (2) The problem is solved into a pair of reasoning complexity when the frequency and the transmitting power are calculated fixedly According to scaling law Total energy consumption Sum end-to-end delay Expressed as: ; ; For a pair of Deriving and zeroing it, the optimal inferred complexity chosen to minimize total energy consumption is recorded as : ; (3) Fixed model complexity and transmit power, calculated frequency Is determined by a delay constraint, which satisfies: ; When selecting frequency When the lower limit is reached, the energy consumption is calculated to be optimal; selecting any one of the conditions (1) - (3), and executing an energy consumption optimization strategy.
  2. 2. The semantic communication method based on deep learning and energy consumption optimization of claim 1, wherein in the step S1, the transmitting end comprises a deep source encoder and a channel encoder, in the deep source encoder, semantic feature extraction and compression are performed on source data by using a deep neural network, and after digital channel coding by the channel encoder, the data are transmitted by using a wireless channel, and the method comprises the following steps: A1, setting a source encoder as a deep neural network source encoder based on a super prior structure, wherein the source encoder comprises a feature extraction network, a quantization module, an entropy encoder, a splicing module and a super prior model, wherein the feature extraction network is a deep neural network; The original source samples of the input are represented as vectors Wherein , Representing the dimensions of the original source data, Representation of The space of the real number is maintained, Obeying statistical distribution ; A2, original source data First, a feature extraction network is input, and a coding function formed by the feature extraction network is used for processing the feature extraction network Mapping to continuous latent variable vectors The relationship is expressed as: ; Wherein the method comprises the steps of Representing a continuous latent variable vector of values, Represents latent variable dimensions and satisfies , A parameter set representing a feature extraction network; a3, continuous latent variable Then the discrete latent variable vector is obtained by the processing of the quantization module Wherein , Representation of A dimension integer space; the quantization module is used for determining continuous latent variables Each element in (2) is rounded to obtain a discrete latent variable vector ; A4, introducing a super prior model to model latent variable distribution: Discrete latent variable Is input into an entropy coder, and the super prior model outputs corresponding distribution parameters according to the statistical characteristics of the latent variables And Wherein A mean vector representing each dimension of the latent variable, Standard deviation vectors representing dimensions of latent variables; The super prior model comprises a super prior encoder, a super prior quantization module and a super prior decoder, wherein the super prior encoder and the super prior decoder are realized through a depth neural network; first, will Inputting the super prior encoder to obtain super prior latent variable : ; Wherein the method comprises the steps of Representing the super a-priori encoder, Representing network parameters of the super a priori encoder; the super prior quantization module performs rounding on Each element of (a) is processed to Quantized to discrete variables ; Will be Inputting the obtained latent variable into a super prior decoder to obtain distribution parameters including average value Sum of variances : ; Wherein the method comprises the steps of Representing the super a-priori decoder, Representing network parameters of the super a priori decoder; Entropy encoder based on distribution parameter pairs Probability modeling and generating bit sequences Simultaneously, the super prior model generates an auxiliary bit sequence for describing the information of the latent variable distribution side ; The super a priori model generates the auxiliary bit sequence by assuming that Obeys a standard Gaussian distribution and is directly to As input, call the entropy encoder to get the bit sequence ; A5, the transmitting end uses the splicing module to splice the bit sequence And (3) with Splicing to form a complete bit stream output by a source encoder And after digital channel coding by a channel coder, the digital channel coding is carried out by utilizing a wireless channel for transmission.
  3. 3. The semantic communication method based on deep learning and energy consumption optimization according to claim 2, wherein the A5 comprises: setting the transmitting end to obtain bit stream after finishing source coding Wherein , Representing the main bit sequence resulting from the entropy encoding, For the super a priori side information bit sequences, And (3) with Respectively represent And (3) with And the total bit length is ; Using a channel encoder to output a compressed bit stream from a source encoder Mapping to complex channel symbol sequences transmitted in a wireless channel, and realizing reliable transmission and delineated shannon transmission rate under given bandwidth and channel conditions: the channel encoder streams the bit stream Coded as length Complex channel symbol sequence of (2) Wherein C represents the complex number field, and the number of the complex number field, Representing the number of channel symbols to be transmitted; and adopting a unit average power normalization constraint on the channel symbols, namely that the average power of each symbol in the expected sense is 1, wherein the expression is as follows: ; Wherein the method comprises the steps of Representing the mathematical expectation operator, Representing vectors Is a binary norm of (2); Let the communication bandwidth be The communication power of the transmitting end for wireless transmission is as follows The equivalent complex channel coefficient from the transmitting end to the receiving end is The noise is additive circular symmetric complex Gaussian noise and its noise power spectral density is The maximum transmission rate supported by the system under the conditions of communication bandwidth and channel according to shannon channel coding theorem, namely shannon rate is recorded as The expression is: ; Wherein the method comprises the steps of A logarithmic operation with a base of 2 is shown, Representing complex channel coefficients Modulus value of (2), shannon rate For inscribing on a given 、 、 、 Maximum reliable transmission capability of channel coding and transmission under the condition, and average code rate of source coding And determining the subsequent transmission delay relation together.
  4. 4. The semantic communication method based on deep learning and energy consumption optimization according to claim 1, wherein in the step S1, the receiving end comprises a deep source decoder and a channel decoder, and the corresponding source data is recovered by the deep source decoder after the channel decoder is used for channel decoding the received signal, comprising: b1, setting an information source decoder to adopt a deep neural network information source decoder based on a super prior structure, wherein the deep neural network information source decoder comprises a segmentation module, an entropy decoder, a super prior model and a characteristic recovery network; B2, the receiving end firstly uses the channel decoder to carry out channel decoding on the received complex channel symbol sequence to obtain bit stream Then the received bit stream is divided into a plurality of bit streams Splitting to obtain corresponding bit sequence And The channel decoding is to restore the complex channel symbol sequence into bit stream b; B3, super prior model according to bit sequence Restoring latent variable distribution parameters And ; Will be As input, directly invoking the entropy decoder to obtain Handle Input to a super a priori decoder to obtain And : ; Wherein the method comprises the steps of Representing the super a-priori decoder, Representing network parameters of the super a priori decoder; B3, entropy decoder based on And For bit sequences Lossless entropy decoding and recovery of discrete latent variable vectors ; B4, recovered discrete latent variable Then inputting the characteristic recovery network to obtain the reconstructed information source data The relationship is expressed as: ; Wherein the feature recovery network is a deep neural network, and the decoding function of the feature recovery network is as follows , Representing the recovered source data vector, A set of parameters representing a feature recovery network.
  5. 5. The semantic communication method based on deep learning and energy consumption optimization according to claim 1, wherein in the step S1, the deep source encoder and the deep source decoder are required to be pre-trained in the following training modes: (1) To vector original source sample As semantic data to be transmitted, after transmission according to the system of step S1, source data recovered by the receiving end is set as ; Defining the distortion of the source data recovery as: ; Representing the vector modulo length; The loss function of model training is defined as: ; Wherein, the For the set distortion threshold value, Representing the average code rate of source coding; extracting parameter set of network for characteristics by using loss function Parameter set for feature recovery network Network parameters of super a priori encoder Parameters of a super a priori decoder Performing joint updating, wherein the updating method is a gradient descent algorithm; (2) For different original source sample vectors Repeating the step (1) until reaching the training round to obtain the final model parameters 、 、 、 And is applied in the actual communication process.
  6. 6. The semantic communication method based on deep learning and energy consumption optimization according to claim 1, wherein in the step S2, the modeling of the source codec energy consumption comprises: providing that the depth source encoder and the depth source decoder are implemented in the same type of processor, for a single source sample in the processor Performing one-time deep source coding and deep source decoding as one-time reasoning calculation, wherein the number of floating point operations required by one-time reasoning calculation is , wherein, The inference calculation complexity for representing floating point operation is set as the calculation frequency of the processor , wherein, For characterizing the number of floating-point operations performed per second by the processor, the calculation delay is inferred once by a single source sample Expressed as: ; Wherein the method comprises the steps of Representing a computation delay; the power consumption model of the processor is adopted to infer the calculation power of the process Expressed as: ; Wherein the method comprises the steps of Representing the calculated power of the computer, Representing constants associated with the hardware architecture, Representing the calculated frequency; Thus, the computational power consumption of single source sample reasoning Expressed as: ; The average code rate of the source coding is , Complexity of reasoning The power law relation is satisfied: ; Wherein the method comprises the steps of And Representing the parameters of a power law fit, Representing the asymptotic lower limit of the compression ratio, with uniform variables At the same time, the average code rate of source coding is depicted AND-calculating energy consumption "Coupling relationship; for the processing of the depth source encoder and the depth source decoder, By adjusting the number of hidden layer channels of the network To achieve continuous variation, wherein Indicating the number of channels of hidden layers in coding network and decoding network Expressed as: ; Wherein the method comprises the steps of Indicating the floating point number of operations of the module except the hidden layer, A floating point operation number coefficient item representing a single-input single-output channel convolution layer, The term is used for representing that when the number of input and output channels is simultaneously The number of connections and the amount of calculation are two times increased.
  7. 7. The semantic communication method based on deep learning and energy consumption optimization according to claim 1, wherein in the step S2, the communication transmission energy consumption modeling comprises: communication transmission energy consumption refers to energy consumed in transmitting a bit stream compressed by source coding in a wireless channel, and average transmission delay Is determined by the number of compressed bits and the channel transmission rate, and specifically: ; on the basis, the energy consumption in the communication stage is controlled by the transmitting power And communication circuit power Co-determination, communication energy consumption using power consumption model Expressed as: ; Wherein the method comprises the steps of Representing the power consumption of a communication circuit, regarding the power consumption as a constant, and expressing the communication transmission energy consumption as an average code rate of source coding through communication transmission energy consumption modeling Communication power Rate of shannon Is a function of (2).

Description

Semantic communication method based on deep learning and energy consumption optimization Technical Field The invention relates to the field of semantic communication, in particular to a semantic communication method based on deep learning and energy consumption optimization. Background As mobile communication systems evolve towards the sixth generation (6G), the communication systems need to support not only higher data rates and lower latency, but also energy efficient operation on energy-constrained terminals and edge devices. The traditional communication system mainly focuses on spectrum efficiency, optimizes transmission rate through shannon theory, but ignores calculation energy consumption in the process of information source processing and communication. In recent years, semantic communication significantly reduces the number of transmission bits by extracting semantic features related to tasks from raw data using a deep neural network, while ensuring perceptual or task performance. However, most of the existing semantic communication research only focuses on transmission performance or perceived accuracy, and it is generally assumed that energy consumption of source coding and channel transmission is negligible, or that computing energy consumption caused by deep neural network reasoning is ignored only by considering communication energy consumption. This modeling approach is difficult to reflect the energy efficiency characteristics of "compute-communicate" deep coupling in practical systems. Therefore, an energy efficiency modeling and optimizing method capable of simultaneously describing the semantic coding complexity, the wireless transmission resource configuration and the end-to-end time delay constraint of the deep neural network is needed to realize the energy efficiency optimal semantic communication facing the actual system. Disclosure of Invention The invention aims to overcome the defects of the prior art, and provides a semantic communication method based on deep learning and energy consumption optimization, which constructs a unified calculation energy consumption and communication energy consumption model, and jointly optimizes key parameters such as complexity of a semantic source coding model, running frequency of a processor, wireless transmission power and the like on the premise of meeting end-to-end time delay constraint, thereby realizing the minimization of the total energy consumption of a system. The invention aims at realizing the technical scheme that the semantic communication method based on deep learning and energy consumption optimization comprises the following steps: Step S1, constructing a deep neural network semantic communication system based on a super prior structure, wherein the deep neural network semantic communication system comprises a transmitting end and a receiving end; The transmitting end comprises a depth source coder and a channel coder, wherein in the depth source coder, a deep neural network is utilized to extract and compress semantic features of source data, and after digital channel coding of the channel coder, a wireless channel is utilized to transmit; the receiving end comprises a depth information source decoder and a channel decoder, and corresponding information source data is recovered through the depth information source decoder after the channel decoder is utilized to carry out channel decoding on the received signals; S2, carrying out energy consumption modeling on a deep neural network semantic communication system, wherein the energy consumption modeling comprises information source coding and decoding energy consumption modeling and communication transmission energy consumption modeling; and step S3, constructing a joint optimization problem by taking the model reasoning complexity, the calculation frequency and the wireless transmission power as adjustable decision variables, and solving to obtain a communication strategy. The method has the beneficial effects that the method combines semantic feature extraction and compression technology based on the deep neural network and digital channel coding and wireless transmission technology, considers the actual scene of coexistence of computing and communication energy consumption in an end-to-end semantic communication system, constructs a unified computing energy consumption and communication energy consumption model, and jointly optimizes key parameters such as complexity of a semantic source coding model, operation frequency of a processor, wireless transmission power and the like on the premise of meeting end-to-end time delay constraint, thereby realizing the minimization of the total energy consumption of the system. Drawings FIG. 1 is a flow chart of the method of the present invention; FIG. 2 is a schematic diagram of a semantic communication system based on a deep neural network; FIG. 3 is a schematic diagram of a depth source encoder based on a super prior structure; FIG. 4