Semantic communication method based on deep learning and energy consumption optimization

Abstract

The invention discloses a semantic communication method based on deep learning and energy consumption optimization. The method comprises the following steps: S1, constructing a deep neural network semantic communication system based on a super-prior structure; s2, performing energy consumption modeling on the deep neural network semantic communication system, wherein the energy consumption modeling comprises information source encoding and decoding energy consumption modeling and communication transmission energy consumption modeling; and S3, taking the model reasoning complexity, the calculation frequency and the wireless transmission power as adjustable decision variables, constructing a joint optimization problem, and solving to obtain a communication strategy. According to the method, a unified calculation energy consumption and communication energy consumption model is constructed, and key parameters such as semantic source coding model complexity, processor operation frequency and wireless transmission power are jointly optimized on the premise of meeting end-to-end time delay constraints, so that the total energy consumption of the system is minimized.

Description

Technical Field

[0001] This invention relates to the field of semantic communication, and in particular to a semantic communication method based on deep learning and energy optimization. Background Technology

[0002] As mobile communication systems evolve towards sixth generation (6G), these systems not only need to support higher data rates and lower latency, but also require energy-efficient operation on energy-constrained terminals and edge devices. Traditional communication systems primarily focus on spectral efficiency, optimizing transmission rates using Shannon's theory, but neglect the computational energy consumption during source processing and communication.

[0003] In recent years, semantic communication has significantly reduced the number of transmitted bits while maintaining perception or task performance by utilizing deep neural networks to extract task-related semantic features from raw data. However, most existing semantic communication research focuses only on transmission performance or perception accuracy, typically assuming that the energy consumption of source coding and channel transmission is negligible, or only considering communication energy consumption while ignoring the computational energy consumption of deep neural network inference. This modeling approach fails to reflect the energy efficiency characteristics of the deep coupling between computation and communication in real-world systems.

[0004] Therefore, there is an urgent need for an energy efficiency modeling and optimization method that can simultaneously characterize the semantic coding complexity of deep neural networks, wireless transmission resource configuration, and end-to-end latency constraints, so as to achieve energy-efficient semantic communication for practical systems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a semantic communication method based on deep learning and energy consumption optimization. It constructs a unified computational energy consumption and communication energy consumption model, and under the premise of satisfying end-to-end latency constraints, jointly optimizes key parameters such as semantic source coding model complexity, processor operating frequency and wireless transmission power, thereby minimizing the total system energy consumption.

[0006] The objective of this invention is achieved through the following technical solution: a semantic communication method based on deep learning and energy consumption optimization, comprising:

[0007] Step S1: Construct a deep neural network semantic communication system based on a super-prior structure, including a sender and a receiver;

[0008] The transmitting end includes a deep source encoder and a channel encoder. In the deep source encoder, a deep neural network is used to extract and compress semantic features of the source data, and after digital channel encoding by the channel encoder, it is transmitted using a wireless channel.

[0009] The receiving end includes a deep source decoder and a channel decoder. After the channel decoder performs channel decoding on the received signal, the deep source decoder recovers the corresponding source data.

[0010] Step S2: Perform energy consumption modeling for the deep neural network semantic communication system, including source encoding / decoding energy consumption modeling and communication transmission energy consumption modeling;

[0011] Step S3: Using model inference complexity, computation frequency, and wireless transmission power as adjustable decision variables, construct a joint optimization problem and solve it to obtain the communication strategy.

[0012] The beneficial effects of this invention are as follows: This invention combines semantic feature extraction and compression technology based on deep neural networks with digital channel coding and wireless transmission technology. Considering the actual scenario where computing and communication energy consumption coexist in end-to-end semantic communication systems, it constructs a unified computing energy consumption and communication energy consumption model. Under the premise of meeting end-to-end latency constraints, it jointly optimizes key parameters such as semantic source coding model complexity, processor operating frequency, and wireless transmission power, thereby minimizing the total system energy consumption. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method of the present invention;

[0014] Figure 2 This is a schematic diagram of a semantic communication system based on deep neural networks.

[0015] Figure 3 This is a schematic diagram of a deep source encoder based on a priori structure.

[0016] Figure 4 This is a schematic diagram of a deep source decoder based on a super-prior structure.

[0017] Figure 5 This is a schematic diagram illustrating the relationship between total energy consumption and reasoning complexity.

[0018] Figure 6 This is a diagram illustrating the relationship between total latency and inference complexity. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0020] like Figure 1 As shown, a semantic communication method based on deep learning and energy consumption optimization includes:

[0021] Step S1: Construct a deep neural network semantic communication system based on a super-prior structure, including a sender and a receiver;

[0022] like Figure 2 The diagram illustrates the architecture of an energy efficiency-aware semantic communication system that integrates deep learning and digital communication technologies. At the transmitting end, the system utilizes a deep neural network to extract and compress semantic features from the source data, and transmits semantic information through digital channel coding and wireless channels. At the receiving end, based on different user needs, the system performs channel decoding on the received signal and recovers the corresponding semantic information or source data.

[0023] like Figure 3 As shown, in this invention, the source coding at the transmitting end adopts a deep neural network image compression model based on a super-prior structure to extract and compress semantic features from the original source data (in this application, semantic data, such as image data).

[0024] A1. Let the source encoder be a deep neural network source encoder based on a priori structure, including a feature extraction network, a quantization module, an entropy encoder, a splicing module, and a priori model; the feature extraction network is a deep neural network;

[0025] The input raw source samples are represented as vectors. ,in , Indicates the dimension of the original source data. express 3D real space, Follows statistical distribution ;

[0026] A2. Original source data First, the input is fed into the feature extraction network, and then processed by the encoding function composed of the feature extraction network. Mapped to continuous latent variable vectors The relationship is expressed as:

[0027] ;

[0028] in Represents a continuous latent variable vector. Represents the dimension of the latent variable, and satisfies , This represents the set of parameters for the feature extraction network.

[0029] A3. Continuous latent variables The discrete latent variable vector is then obtained after processing by the quantization module. ,in , express 3D integer space;

[0030] The quantization module performs analysis on continuous latent variables. Each element in the vector is rounded to obtain the discrete latent variable vector. ;

[0031] A4. Introduce a prior model to model the distribution of latent variables:

[0032] Discrete latent variables The input is given to the entropy encoder, and the prior model outputs the corresponding distribution parameters based on the statistical properties of the latent variables. and ,in This represents the mean vector of each dimension of the latent variable. This represents the standard deviation vector of each dimension of the latent variable;

[0033] The super-prior model includes a super-prior encoder, a super-prior quantization module, and a super-prior decoder; the super-prior encoder and super-prior decoder are both implemented using deep neural networks.

[0034] First, Input the hyper-prior encoder to obtain hyper-prior latent variables :

[0035] ;

[0036] in Indicates a priori encoder, This represents the network parameters of the hyper-prior encoder;

[0037] The prior quantization module uses rounding to... Process each element in the process, and Quantized into discrete variables ;

[0038] Will The distribution parameters of the latent variables, including the mean, are obtained by inputting the super-prior decoder. and variance :

[0039] ;

[0040] in Indicates a priori decoder, This represents the network parameters of the hyper-prior decoder;

[0041] Entropy encoders are based on distributed parameter pairs Perform probabilistic modeling and generate bit sequences Simultaneously, the advanced prior model generates an auxiliary bit sequence to describe the information on the distribution side of the latent variables. ;

[0042] The advanced prior model generates the auxiliary bit sequence in the following way: assuming Following a standard Gaussian distribution, directly... As input, the entropy encoder is called to obtain the bit sequence. ;

[0043] A5. The transmitting end uses a bit sequence splicing module to... and The bits are spliced ​​together to form the complete bit stream output by the source encoder. After digital channel encoding via a channel encoder, it is transmitted using a wireless channel.

[0044] like Figure 4 As shown, at the receiving end, the source decoding module is used to recover the latent variable representation from the received bit stream and reconstruct the source data.

[0045] B1. Suppose that the source decoder adopts a deep neural network source decoder based on a priori structure, including a segmentation module, an entropy decoder, a priori model, and a feature recovery network (also called an image recovery network when the source data is an image).

[0046] B2. The receiver first uses a channel decoder to perform channel decoding on the received complex channel symbol sequence to obtain the bit stream. Then, the segmentation module is used to process the received bit stream. Decompose the data to obtain the corresponding bit sequence. and The channel decoding process involves restoring a complex sequence of channel symbols to a bit stream b.

[0047] B3. The advanced prior model is based on the bit sequence. Recovery of latent variable distribution parameters and ;

[0048] Will As input, the entropy decoder is called directly to obtain... ,Bundle The input is fed into the hyper-prior decoder to obtain and :

[0049] ;

[0050] in Indicates a priori decoder, This represents the network parameters of the hyper-prior decoder;

[0051] B3, Entropy Decoder Based on and Bit sequence Perform lossless entropy decoding to recover the discrete latent variable vector. ;

[0052] B4. Recovered discrete latent variables The data is then input into the feature recovery network to obtain the reconstructed source data. The relationship is expressed as:

[0053] ;

[0054] The feature recovery network is a deep neural network, and its decoding function is: , This represents the recovered source data vector. This represents the parameter set of the feature recovery network. Through the above process, the receiver can achieve high-quality reconstruction of the original source data while ensuring compression efficiency.

[0055] In this invention, the channel coding module is used to compress the bit stream output by the source coding module. This is further mapped to a complex sequence of channel symbols that can be transmitted over a wireless channel, in order to achieve reliable transmission under given bandwidth and channel conditions and characterize the achievable Shannon transmission rate. Specifically, the transmitter obtains the bit stream after completing source coding. ,in , This represents the main bit sequence obtained by entropy encoding. This represents the sequence of bits representing the prior information. and These represent the lengths of the corresponding bit sequences, and the total bit length is... The channel encoder will output the bit stream. Encoded as a length of Complex channel symbol sequence ,in C represents the complex field. This represents the number of channel symbols to be transmitted. For ease of unified modeling, this invention employs a unit average power normalization constraint for the channel symbols, meaning that the average power of each symbol is expected to be 1, expressed as:

[0056] ;

[0057] in Represents the mathematical expectation operator. Representing vectors The norm of 2. Combining this with the subsequent wireless transmission model, let the communication bandwidth be... (Unit: Hz), the communication power used by the transmitting end for wireless transmission is (Unit: W), the equivalent complex channel coefficients from the transmitter to the receiver are: The noise is additive circularly symmetric complex Gaussian noise with a noise power spectral density of... (Unit: W / Hz), then according to Shannon's channel coding theorem, the maximum transmission rate (Shannon rate) that the system can support under the above bandwidth and channel conditions is denoted as: (Unit: bit / s), its expression is:

[0058] ;

[0059] in This represents the logarithmic operation with base 2. Represents complex channel coefficients The modulus; the Shannon rate Used to characterize in a given , , , Under the given conditions, the maximum reliable transmission capability achievable through channel coding and transmission, and the average code rate of the aforementioned source coding, are compared. Together they determine the subsequent transmission delay relationship, thus providing a unified parameter interface for the joint configuration of power and rate in energy efficiency optimization.

[0060] In step S1, the deep source encoder and deep source decoder need to be pre-trained. The training method is as follows:

[0061] (1) Convert the original source sample vector As semantic data to be transmitted, after being transmitted according to the system in step S1, let the source data recovered by the receiving end be... ;

[0062] The distortion in source data recovery is defined as:

[0063] ;

[0064] Indicates the magnitude of the vector;

[0065] The loss function for model training is defined as:

[0066] ;

[0067] in, The set distortion threshold, This represents the average code rate of the source coding;

[0068] Using a loss function, the parameter set of the feature extraction network is... The parameter set of the feature recovery network Network parameters of the priori encoder Parameters of the super-prior decoder A joint update is performed using the gradient descent algorithm.

[0069] (2) For different original source sample vectors Repeat step (1) until the training rounds are reached, and then obtain the final model parameters. , , , And apply it in actual communication processes;

[0070] In the embodiments of this application, during each training iteration, the average code rate of the source coding will change with the model parameters. , , , The model changes with updates; after training, the model parameters are fixed. It also remains unchanged.

[0071] Step S2: Perform energy consumption modeling for the deep neural network semantic communication system, including source encoding / decoding energy consumption modeling and communication transmission energy consumption modeling;

[0072] In this invention, the depth source encoder and depth source decoder are implemented in the same type of processor; the processor processes individual source samples. One deep source coding and one deep source decoding operation is denoted as one inference computation. The number of floating-point operations required for one inference computation is: ,in, Used to characterize the inference computational complexity of floating-point operations; assuming the processor's computational frequency is... ,in, The computational latency for a single inference operation on a single source sample is used to characterize the number of floating-point operations performed by the processor per second. Represented as:

[0073] ;

[0074] in Indicates computation delay;

[0075] Using a processor power consumption model, the computational power of the inference process is... Represented as:

[0076] ;

[0077] in Indicates calculated power. This represents a constant related to the hardware architecture. Indicates the calculated frequency;

[0078] Therefore, the computational energy consumption of inference from a single source sample is... Represented as:

[0079] ;

[0080] The average code rate of the source coding is denoted as . , with reasoning complexity They satisfy a power-law relationship:

[0081] ;

[0082] in and Represents the power-law fitting parameters. The asymptotic lower bound of the compression ratio is represented by a uniform variable. Simultaneously characterize the "average code rate of the source coding" "and "calculate energy consumption" The coupling relationship between "".

[0083] The processing steps for depth source encoders and depth source decoders, By adjusting the number of hidden layer channels in the network To achieve continuous change, where This represents the number of channels in the hidden layers of the encoding and decoding networks; at this time Represented as:

[0084] ;

[0085] in This indicates the number of floating-point operations performed by modules other than the hidden layer. This represents the floating-point operation coefficients for a single-input, single-output channel convolutional layer. This item is used to indicate when both the number of input and output channels are 0. The number of connections and computational cost exhibit a quadratic growth characteristic.

[0086] Through the above modeling, this invention directly uses inference complexity without introducing block coding assumptions. With calculation frequency To quantify the energy consumption of source coding and the average code rate of the aforementioned source coding Maintaining consistent notation provides a resolvable energy consumption expression for subsequent joint "computation-communication" energy efficiency optimization.

[0087] In this invention, communication transmission energy consumption refers to the energy consumed when transmitting a source-coded and compressed bit stream in a wireless channel. Average transmission delay. (Unit: s) is determined by the number of compressed bits and the channel transmission rate, specifically:

[0088] ;

[0089] Based on this, the energy consumption during the communication phase is determined by the transmission power. and communication circuit power It was jointly decided that a common power consumption model would be adopted for communication energy consumption. (Unit: J) is expressed as:

[0090] ;

[0091] in This indicates the power consumption of the communication circuit, mainly generated by digital signal processing modules, mixers, low-noise amplifiers, and synchronizers; for systems with fixed bandwidth and data rates, The variation is small and can be considered a constant. Through the above modeling, this invention expresses the communication transmission energy consumption as the average code rate of the source coding, while maintaining symbolic consistency with the energy consumption modeling of source coding. Communication power and Shannon rate The function provides a clear parameter interface and energy consumption expression for subsequent joint optimization of computation and communication energy efficiency.

[0092] Step S3: Using model inference complexity, computation frequency, and wireless transmission power as adjustable decision variables, construct a joint optimization problem and solve it to obtain the communication strategy.

[0093] To leverage the aforementioned scaling law to achieve high energy efficiency in semantic communication systems, this invention uses model inference complexity, computation frequency, and wireless transmission power as adjustable decision variables to construct a joint optimization problem. Total energy consumption End-to-end delay They are represented as follows:

[0094] ;

[0095] ;

[0096] Given maximum allowable end-to-end delay The optimization objective of this invention is to minimize total energy consumption while satisfying time delay constraints. The optimization problem can be expressed as:

[0097] ;

[0098] This problem involves finding decision variables. , and The joint configuration makes the objective function Taking the minimum value, the constraint means that the end-to-end delay does not exceed a given threshold. In this problem, Control source encoding output bit count and computational delay ; Determines the latency and energy consumption of inference calculations; Simultaneously affects Shannon's rate and communication energy consumption ; in constraints Given the application scenario, this represents the user experience or system real-time requirements. Through this modeling, guided by the scaling law, this invention couples the configuration of computing and communication resources into a unified energy-latency optimization framework.

[0099] To solve the aforementioned joint optimization problem, this invention first fixes some of the variables, gradually analyzes the coupling relationship between the decision variables, and then uses scaling laws and energy efficiency models to derive a closed-form solution.

[0100] (1) Fixed model complexity and calculate frequency At that time, the problem degenerated into concerns about the transmission power. The univariate optimization aims to minimize communication energy consumption. Under a given time delay constraint, this is equivalent to maximizing energy efficiency. Among them, energy efficiency Defined as:

[0101] ;

[0102] in Modulo h; energy efficiency That is, the number of effective bits that can be transmitted per unit of energy; by... for By taking the derivative and making it zero, the optimal transmission power is obtained;

[0103] (2) When the calculation frequency and transmission power are fixed, the problem degenerates into a problem of reasoning complexity. Optimization, based on scaling laws Total energy consumption and end-to-end delay They are represented as follows:

[0104] ;

[0105] ;

[0106] Taking the derivative of the above expression and setting it to zero, we obtain the optimal reasoning complexity chosen to minimize the total energy consumption, denoted as . Its expression is:

[0107] ;

[0108] (3) With fixed model complexity and transmission power, calculate the frequency. The lower bound is determined by the time delay constraint, which satisfies:

[0109] ;

[0110] When selecting frequency When the lower bound is reached, the computational energy consumption is minimized;

[0111] S303. Select any one of the cases (1) to (3) in step S302 and execute the energy-optimal strategy.

[0112] In the embodiments of this application, the application is verified through simulation experiments:

[0113] A simulation platform was built and its parameters were set. Firstly, on the computing side, an AMD EPYC8534PN CPU, designed for edge inference, was used for deep neural network model inference. This chip is based on the Zen4c microarchitecture and supports the AVX-512 instruction set; its base clock frequency is [missing information]. =2.0GHz. The single-core computing frequency ν (unit: FLOP / s) is determined by the processor's clock frequency and vector instruction parallelism, and the calculation expression is:

[0114] ;

[0115] in This indicates the number of data elements that a single vector instruction can process simultaneously. This indicates the number of parallel units in the processor that support AVX-512 instructions, with a coefficient of 2 introduced due to the use of hybrid multiply-accumulate (FMA) operations. For the AMD EPYC8534PN, the maximum single-core computing frequency is... =128 GFLOPs / s. To simplify the model and facilitate energy consumption calculation, a calculation energy efficiency constant is introduced. This is used to map computation frequency to power consumption; based on processor specifications and existing models. The relationship between source coding rate and inference complexity is parameterized using a power-law model based on scaling laws, and the power-law coefficients are obtained through fitting. =4.63×10^{12}、 =1.77 and the lower bound of source entropy =0.212.

[0116] On the communication side, the simulation uses a large-scale path loss channel model with a path loss exponent. =3.5, communication bandwidth =1MHz, noise power spectral density =−174dBm / Hz. Considering power amplifier efficiency. =0.35, transmit antenna gain =5dBi, receiver antenna gain =0dBi, Noise Figure =10dB, link margin compensation =10dB. The system operates in the 3.5GHz band, and the link signal-to-noise ratio is... The calculation formula is:

[0117] ;

[0118] in The distance between the transmitter and receiver is set in this experiment. =200 meters, =32.44dB is the reference gain factor. The power consumption of the communication circuit is taken as a constant. =2W. To test the energy efficiency of the semantic communication link, color images from the CIFAR-10 dataset were selected on the data side, with 64×64 RGB images used as the source input.

[0119] Simulations were performed based on the above settings to compare the total energy consumption and end-to-end latency of the system under different inference complexity and transmit power configurations, and the results were compared with the prediction results of the scaling law. The experiment first fixed the transmit power and changed the number of floating-point operations in the neural network inference. Calculate total energy consumption and latency The changes. Figure 5 and Figure 6 Simulation results show that total energy consumption and end-to-end delay increase with... The changes show a trend of first decreasing and then increasing: when the inference complexity is low, the source coding rate is high and the transmission latency is large, resulting in high total energy consumption; as... Increasing the source coding rate reduces the transmission delay and decreases the total energy consumption; when As it continues to increase, calculate the energy consumption factor. The rapid increase in relationships leads to a rise in total energy consumption. During this process, an optimal reasoning complexity can be observed. This minimizes the total energy consumption, a value consistent with the aforementioned analytical expression; there is also an optimal reasoning complexity for minimizing end-to-end latency. The lowest point of the time delay curve is reflected in the simulation curve.

[0120] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A semantic communication method based on deep learning and energy optimization, characterized in that: The method comprises the following steps: Step S1: constructing a deep neural network semantic communication system based on a hyper-prior structure, including a sending end and a receiving end; The sending 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, wireless channel transmission is performed; The receiving end comprises a deep source decoder and a channel decoder, after channel decoding of the received signal by the channel decoder, the corresponding source data is recovered by the deep source decoder; Step S2: energy consumption modeling for the deep neural network semantic communication system, including source coding and decoding energy consumption modeling and communication transmission energy consumption modeling; Step S3: taking model inference complexity, calculation frequency and wireless transmission power as adjustable decision variables, constructing a joint optimization problem, and solving to obtain a communication strategy.

2. The semantic communication method based on deep learning and energy optimization according to claim 1, characterized in that: In step S1, the sending 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, wireless channel transmission is performed, comprising: A1, the source encoder is a deep neural network source encoder based on a hyper-prior structure, comprising a feature extraction network, a quantization module, an entropy encoder, a splicing module and a hyper-prior model; the feature extraction network is a deep neural network; The input raw source samples are represented as a vector wherein , denotes the dimension of the raw source data, denotes a d-dimensional real space, obeys a statistical distribution ; A2, raw source data First, the feature extraction network is inputted, and an encoding function constituted by the feature extraction network is applied is mapped to a continuous latent variable vector The relationship is expressed as: ； wherein represents a continuous latent variable vector, represents a latent variable dimension, and satisfies , represents a parameter set of the feature extraction network; A3, continuous latent variable Subsequently, the discrete latent variable vector is obtained by quantization module processing wherein , denotes dimensional integer space; The quantization module obtains a discrete latent variable vector by rounding each element in the continuous latent variable vector ​ A4, a hyper-prior model is introduced to model the latent variable distribution: discrete latent variables are input to the entropy encoder, while the hyper-prior model outputs corresponding distribution parameters according to the latent variable statistical properties and where denotes the mean vector of each dimension of the latent variable, denotes the standard deviation vector of each dimension of the latent variable; Wherein, the hyper-prior model comprises a hyper-prior encoder, a hyper-prior quantization module and a hyper-prior decoder; wherein the hyper-prior encoder and the hyper-prior decoder are realized by a deep neural network; First, the input is encoded by the hyper-prior encoder to obtain the hyper-prior latent variable input hyper-prior encoder to obtain a hyper-prior latent variable : ； wherein represents a hyper-prior encoder, represents network parameters of the hyper-prior encoder; The super-prior quantization module processes each element in the matrix by rounding off to obtain a discrete variable ;​​ will be described in detail below. The input hyper-prior decoder obtains the distribution parameters of the latent variables, including the mean and variance : ； wherein denotes a hyper-prior decoder, denotes a network parameter of the hyper-prior decoder; An entropy encoder models probabilities based on distribution parameters and generates bit sequences ; meanwhile, a hyper-prior model generates auxiliary bit sequences for describing side information of the latent variable distribution ; The super-prior model generates the auxiliary bit sequence in the following manner: assuming subject to a standard Gaussian distribution, directly inputting into the entropy encoder to obtain a bit sequence ; A5, the transmitting end splices the bit sequence with to form a complete bit stream output by the source encoder and after digital channel coding by the channel encoder, transmits by using the wireless channel.

3. The semantic communication method based on deep learning and energy optimization according to claim 2, characterized in that: Said A5 comprises: Let the transmitter obtain a bit stream after completing source coding wherein , denotes a main bit sequence obtained by entropy coding, is a hyper-prior side information bit sequence, and denote bit sequence lengths of and respectively, and the total bit length is ; Using a channel encoder, the compressed bitstream output by the source encoder is... The Shannon transmission rate, mapped as a sequence of complex channel symbols transmitted in a wireless channel, achieves reliable transmission under given bandwidth and channel conditions: A channel encoder encodes a bit stream into a sequence of complex channel symbols of length where C denotes the complex number field, Nc denotes the number of channel symbols to be transmitted.​ The channel symbol adopts unit average power normalization constraint, that is, the average power of each symbol in the expectation is 1, and its expression is: ； wherein denotes the mathematical expectation operator, denotes the two-norm of the vector ; Let the communication bandwidth be , the communication power for wireless transmission of the sending end be , the equivalent complex channel coefficient from the sending end to the receiving end be , and the noise be additive circularly symmetric complex Gaussian noise with a noise power spectral density of , then according to the Shannon channel coding theorem, the maximum transmission rate supported by the system under the communication bandwidth and channel conditions, i.e. the Shannon rate, is denoted as , and the expression is: ； wherein denotes the logarithm to the base 2, denotes the complex channel coefficient of the module; the Shannon rate is used to describe the maximum reliable transmission capacity of channel coding and transmission under the given , , , condition, and together with the average code rate of the source coding determines the subsequent transmission delay relationship.

4. The semantic communication method based on deep learning and energy optimization according to claim 1, characterized in that: In step S1, the receiving end comprises a deep source decoder and a channel decoder, after channel decoding of the received signal by the channel decoder, the corresponding source data is recovered by the deep source decoder, comprising: B1, the source decoder adopts a deep neural network source decoder based on a hyper-prior structure, comprising a segmentation module, an entropy decoder, a hyper-prior model and a feature recovery network; B2, the receiving end first utilizes a channel decoder to perform channel decoding on the received complex channel symbol sequence to obtain a bit stream ; then utilizes a splitting module to split the received bit stream to obtain corresponding bit sequences and ; the channel decoding is to restore the complex channel symbol sequence to the bit stream b; B3, the super-prior model is based on the bit sequence recovering the latent variable distribution parameters and ; Call the entropy decoder directly as input to get As input, call the entropy decoder directly to get Call the hyper-prior decoder to get and and : ； wherein represents a hyper-prior decoder, represents a network parameter of the hyper-prior decoder; B3. entropy decoder based on and to the bit sequence performing lossless entropy decoding, recovering the discrete latent variable vector ; B4, recovered discrete latent variable subsequently input into a feature recovery network to obtain reconstructed 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 , denotes the recovered source data vector, denotes a set of parameters of the feature recovery network.

5. The semantic communication method based on deep learning and energy optimization according to claim 1, characterized in that: In step S1, the deep source encoder and the deep source decoder need to be pre-trained, and the training method is: (1) the original source sample vector As the semantic data to be transmitted, the system according to step S1 transmits, and supposes that the source data recovered by the receiving end is ; The distortion of source data recovery is defined as: ； denotes the vector norm; The loss function of model training is defined as: ； wherein is a set distortion threshold, denotes the average code rate of the source coding; a parameter set of the feature extraction network using a loss function a parameter set of the feature recovery network a network parameter of the hyper-prior encoder a parameter of the hyper-prior decoder perform joint updating, and the updating method is a gradient descent algorithm (2) For different original source sample vectors , repeat (1) until the training round is reached, and obtain the final model parameters 、 、 、 , and apply in actual communication process.

6. The semantic communication method based on deep learning and energy optimization according to claim 1, characterized in that: In step S2, the source coding and decoding energy consumption modeling comprises: Let the depth source encoder and the depth source decoder be implemented in the same type of processor; in the processor, a single source sample Performing one depth source encoding and depth source decoding is called one inference calculation, and the number of floating point operations required for one inference calculation is , wherein The inference calculation complexity is used to characterize the floating point operation; let the calculation frequency of the processor be , wherein The number of floating point operations completed per second by the processor is used to characterize the processor, and then the calculation delay of a single source sample once inference is It is expressed as: ； wherein represents the computation latency; Using a processor power model, the computational power of the inference process is is expressed as: ； wherein represents the computing power, represents a constant related to the hardware architecture, represents the computing frequency; Thus, the computational energy consumption of individual source sample inference is represented as: ； The average code rate of the source coding is , and the inference complexity satisfy a power-law relationship: ； wherein and denotes the power law fitting parameter, denotes the asymptotic lower bound of the compression rate, in terms of the unified variable simultaneously characterizes the coupling relationship between the "average code rate of source coding " and the "computational energy cost ". For the processing procedure of the depth signal source encoder and the depth signal source decoder, By adjusting the number of channels of the hidden layer of the network To realize continuous change, wherein Indicates the number of channels of the hidden layer in the encoding network and the decoding network; at this time Indicates that: ； wherein represents the number of floating point operations of the module except the hidden layer, represents the number of floating point operations coefficient term of single input single output channel convolution layer, the term is used to reflect the characteristics of quadratic growth of connection number and calculation amount when the input and output channel number are ​ 7. The semantic communication method based on deep learning and energy optimization according to claim 1, characterized in that: In step S2, the communication transmission energy consumption modeling comprises: The communication transmission energy consumption refers to the energy consumed in transmitting the source coding compressed bit stream in a wireless channel, and the average transmission delay The communication transmission energy consumption is determined by the compressed bit number and the channel transmission rate, and is specifically as follows: ； On this basis, the energy consumption of the communication stage is jointly determined by the transmission power and the communication circuit power and is expressed by using the power consumption model as follows: ​ ； wherein represents the communication circuit power consumption, considered constant, modeling the communication transmission energy consumption as a function of the average code rate of the source coding , of the communication power and of the Shannon rate . 8.The semantic communication method based on deep learning and energy consumption optimization of claim 1, wherein: The step S3 comprises: S301. Taking model inference complexity, calculation frequency and wireless transmission power as adjustable decision variables, constructing a joint optimization problem: Total energy consumption With end-to-end latency Respectively ； ； Given the maximum allowed end-to-end latency The optimization objective is to minimize the total energy consumption while satisfying the latency constraint, and the optimization problem is formulated as: ； The problem finds the joint configuration of decision variables , and to minimize the objective function subject to constraints that represent end-to-end latency not exceeding a given threshold; in the optimization problem, controls the number of source encoding output bits and computes the latency ; decides the inference computation latency and energy consumption; simultaneously affects the Shannon rate and the communication energy ; in the constraints, represents the user experience or system real-time requirement; S302. Fixing part of the variables, gradually analyzing the coupling relationship between the decision variables, and then deriving a closed-form solution by using the scaling law and the energy efficiency model: (1) Fixed model complexity and computation frequency , the problem reduces to a single-variable optimization of the transmit power whose objective is to minimize the communication energy cost , which is equivalent to maximizing the energy efficiency under a given latency constraint where the energy efficiency is defined as: ； wherein is the modulo of h; energy efficiency i.e. the number of effective bits that can be transmitted per unit of energy; by dividing for derivation and making the derivative zero, the optimal transmission power is obtained; (2) When the computing frequency and the transmission power are fixed, the problem degenerates into the optimization of the inference complexity , according to the scaling law , the total energy consumption and the end-to-end latency are respectively represented as: ； ； To derivative and set it to zero, we obtain the optimal inference complexity chosen to minimize the total energy consumption, denoted as : ； (3) Fixed model complexity and transmit power, computation frequency The lower bound of the frequency is determined by the latency constraint, which satisfies: ； When selecting the lower bound of the frequency the energy consumption is calculated to be optimal; S303. Selecting any one of (1) to (3) in step S302, and executing the energy optimal strategy.

Images