Semantic communication cross-layer design method based on intelligent reflection surface

Abstract

The invention relates to a semantic communication cross-layer design method based on an intelligent reflection surface, and belongs to the technical field of communication. In order to solve the problem that robustness and priority processing of a traditional communication system in semantic information transmission are limited, semantic importance and semantic robustness in text data are extracted through a semantic layer, and the priority of data streams is determined on the basis of a two-dimensional feature weighting model. And then, a channel is optimized by using an intelligent reflection surface IRS, and a propagation path of a signal is adjusted in a physical layer, so that the transmission effect of a high-priority data stream is improved. By alternately optimizing beam forming precoding and IRS reflection phase shift, the receiving signal power is maximized, and the interference of noise on low-priority data streams is reduced. According to the method, the anti-interference capability and the transmission efficiency of a semantic communication system are remarkably improved, key semantic features can be protected preferentially in a complex channel environment, and efficient and robust transmission of semantic information is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of communications and relates to a semantic communication cross-layer design method based on intelligent reflective surfaces. Background Art

[0002] With the widespread ubiquitous access of smart terminals and the exponential growth of wireless data traffic, 6G communication systems face severe challenges in terms of spectrum resource utilization efficiency and communication transmission reliability. Against this backdrop, semantic communication, by constructing a semantic parsing framework for joint source-channel coding, deeply explores the semantic representation and contextual association characteristics of information transmission. This significantly reduces bandwidth requirements and greatly improves communication efficiency without compromising user experience, providing a new theoretical framework and technical path for the development of future intelligent communication systems. Currently, the deep integration of deep learning and semantic communication has become an inevitable trend. Most semantic communication systems optimize the transmission process from sender to receiver through deep learning-driven joint source-channel coding technology and end-to-end training, achieving accurate and efficient communication.

[0003] Semantic features are the core elements in semantic communication systems. They directly determine the semantic expression and transmission efficiency of information. With the deep integration of natural language processing, computer vision and semantic communication, the logical completeness and interpretability of semantic feature extraction will be further improved, which will greatly contribute to the performance improvement of semantic communication systems. The key semantic features frequently involved in current research are semantic importance and semantic robustness: semantic importance is reflected in its contribution to overall semantic understanding, while semantic robustness reflects its ability to resist interference under noise interference. In complex channel environments, how to effectively identify and prioritize the transmission of semantic features with high semantic importance and low robustness is the key to improving the performance of semantic communication systems. How to combine these two key features to capture the deep meaning in text or images still requires further research.

[0004] Existing research is mostly based on simplified channel assumptions and cannot effectively address the impact of actual channel impairments such as multipath fading on semantic recovery. In this context, intelligent reflective surfaces (IRS), also one of the key enabling technologies for 6G, have attracted widespread attention due to their ability to dynamically optimize the channel environment. IRS consists of a large number of low-cost passive reflective elements that independently adjust the phase shift and amplitude of the incident electromagnetic wave in a programmable manner, providing an effective solution for physical layer optimization of semantic communication systems in complex channel environments. However, how to further integrate the dynamic optimization of the physical layer IRS with the semantic information transmission in semantic communication remains an urgent issue for researchers.

[0005] Current research has made some progress in extracting semantic features or improving channel transmission performance, but it is limited to physical layer channel optimization or semantic layer feature extraction, and has failed to establish a cross-layer mapping mechanism between semantic features and physical channels. Therefore, the coordination mechanism between the physical layer and the semantic layer urgently needs in-depth research. Summary of the Invention

[0006] In view of this, in order to solve the problems of the above-mentioned prior art and realize cross-layer mapping of semantic features and physical channels, the present invention performs cross-layer optimization design from two aspects: the semantic layer, features of semantic importance and semantic robustness are extracted, and the priority order of text transmission is determined based on two-dimensional feature weighting; the physical layer, a closed solution for maximizing the transmission power is obtained by alternately optimizing the active precoding beamforming vector and the IRS passive phase shift; a cross-layer design framework is proposed, which uses IRS-assisted fading channels to transmit high-priority data streams, and uses non-IRS-assisted fading channels to transmit low-priority data streams.

[0007] A cross-layer design method for semantic communication based on intelligent reflective surfaces includes the following steps:

[0008] S1: Use the Bidirectional Encoder Representations from Transformers (BERT) pre-trained model and combine it with cosine similarity to evaluate the semantic relevance loss of each word to measure the semantic importance of the transmitted text;

[0009] S2: Based on the mean squared error (MSE) of the incremental loss generated by the semantic communication system before and after the feature vector perturbation, a mapping of the word's ability to suppress channel noise distortion is established, thereby characterizing the semantic robustness of the transmitted text;

[0010] S3: Based on the weighted combination of two-dimensional features, a joint measurement model is constructed to determine the feature priority of the transmitted text and divide it into high-priority data streams and low-priority data streams;

[0011] S4: A joint function of the physical layer and the semantic layer is introduced to characterize the matching degree between the recovered estimated text and the input text, and it is decoupled into a channel function that maximizes signal power and a network function that minimizes cross entropy loss;

[0012] S5: Use an alternating optimization algorithm to maximize the received signal power and obtain the optimal dynamic beamforming precoding vector and IRS passive phase shift coefficient matrix;

[0013] S6: With the goal of minimizing semantic information errors, the stochastic gradient descent method is used to minimize cross-entropy loss. The weights and biases of the semantic neural network are repeatedly trained to enhance the network's anti-distortion capabilities. Based on the optimized parameters, the text semantic information is transmitted.

[0014] S7: Transmit text semantic information based on the feature priority list index and the optimized IRS reflection phase shift coefficient matrix: Use IRS to transmit high-priority data streams and fading channels to transmit low-priority data streams, thereby achieving favorable protection of key semantic features and transmission of semantic information.

[0015] Furthermore, the S1 comprises the following steps:

[0016] The preprocessed text s is fed into the pre-trained language model BERT, and the vector representation B of each word based on the BERT model is extracted. ψ (·), then based on the extracted vector representation B ψ (·) Use cosine similarity to calculate the semantic relevance between two words in the lth sentence:

[0017]

[0018] Where i and j represent the i-th and j-th words in the sentence respectively, the superscript T represents the matrix transpose, ||·|| represents the L1 norm, ψ(s l,i ,s l,j ) represents the semantic relevance between the i-th and j-th words in the l-th sentence;

[0019] In order to accurately quantify the semantic contribution of a word in the context, the average semantic association between the target word and all other words in the sentence is established as follows:

[0020]

[0021] Among them, d represents the length of the sentence, Ψ(s l,i ) represents the average semantic relevance of the i-th word in the l-th sentence;

[0022] Based on the above semantic relevance calculation method, the semantic relevance loss caused by a word in a sentence is used to define its semantic importance: when a word has a significant semantic relevance to the context words, removing the word will lead to a large semantic change, which indicates that the word has a high semantic contribution in the semantic composition of the sentence. For the i-th word in the l-th sentence in s, its semantic importance score I(s) is l,i ) is expressed as:

[0023] I(s l,i )=1-Ψ(s l,i ).

[0024] Furthermore, the step S2 includes the following steps:

[0025] The preprocessed text s is passed through the semantic encoder Then it is mapped into a high-dimensional feature vector

[0026] right After applying additive perturbation, the i-th feature dimension of the perturbation feature vector obtained after the channel codec is expressed as

[0027] The loss increment of each feature dimension directly reflects the noise tolerance of different semantic units, and thus characterizes their semantic robustness level. The loss increment is measured by MSE:

[0028]

[0029] Among them, n represents the feature dimension size mapped after the semantic encoder, Γ k represents the loss increment of the k-th word;

[0030] If the loss increment Γ of the feature vector corresponding to the word k The larger the value, the worse its semantic robustness is. Then the semantic robustness score r of the kth word is k Defined as Γ k The reciprocal of:

[0031]

[0032] Furthermore, the step S3 includes the following steps:

[0033] σ k =λ·I k +μ·(1-r k )

[0034] Among them, I k and r k denote the semantic importance score and semantic robustness score of the kth word, λ and μ denote the system’s preference coefficients for semantic importance and semantic robustness, and σ k represents the feature priority score of the kth word in the text;

[0035] Finally, the feature priority scores of each word are arranged in descending order and recorded as σ=[σ1,σ2,…,σ c ], and send σ into the background knowledge base. During transmission, the text data stream will be divided into high-priority data stream and low-priority data stream according to the feature priority score list.

[0036] Furthermore, in S4, in order to characterize the matching degree between the estimated text after the system is restored and the input text, a joint function combining two different layers, the physical layer and the semantic layer, is introduced and named the semantic matching function Mat(·):

[0037]

[0038] stC1:||w t || 2 ≤P max ,

[0039]

[0040] C3:β,α,χ,δ are all hyperparameters

[0041] Among them, w t represents the beamforming precoding vector, Φ represents the IRS reflection phase shift coefficient matrix, γ R represents the signal power, Γ CE represents the cross entropy loss, P max represents the maximum power constraint, n represents the nth IRS element, θ n represents the phase shift of the nth element in the IRS reflection phase shift coefficient matrix, C1 represents constraint 1, C2 represents constraint 2, and C3 represents constraint 3;

[0042] By maximizing the received signal power and minimizing the cross-entropy loss, the matching degree between the input text and the estimated text is maximized, and then it is decoupled into a channel function that maximizes the signal power and a network function that minimizes the cross-entropy loss.

[0043] Furthermore, the corresponding sub-problem in the physical layer of the joint function is decoupled into a channel function that maximizes the signal power and solved separately. It can be expressed as: enhancing wireless transmission on the physical channel requires the joint design of the active beamforming precoding vector w t and the passive phase shift coefficient θ of the IRS n , to maximize the received signal power γ R :

[0044]

[0045] stC1:||w t || 2 ≤P max ,

[0046]

[0047] Among them, |·| represents the absolute value, H dr ∈C N×1 Indicates the direct link from the transmitter to the receiver, HT ∈C M×N and H R ∈C M×1 Denote the equivalent baseband channels from the transmitter to the IRS and from the IRS to the receiver, respectively. The set of reflection elements of the IRS is represented by M. For each element β m ∈[0,1] and θ m ∈[0,2π) represent the amplitude and phase shift of the mth element of IRS respectively. The reflection coefficient matrix of IRS is expressed as

[0048] For the above problem, an alternating optimization algorithm is used to solve it. The transmit beamforming vector at the transmitter and the phase shift at the IRS are iteratively optimized in an alternating manner. One is fixed each time until the two converge or the maximum number of iterations is reached, and the optimal transmit beamforming precoding vector is obtained. and the reflection phase shift coefficient matrix Φ * .

[0049] Furthermore, the corresponding sub-problems in the semantic layer of the joint function are decoupled into network functions that minimize the cross entropy loss and solved separately:

[0050]

[0051] stC3:β,α,χ,δ are all hyperparameters

[0052] This sub-problem aims to minimize the semantic information error and uses the cross entropy function as the loss function to quantify the difference between the input text sequence and the output text sequence;

[0053] First, the loss of the receiving end is calculated and back-propagated to the sending end. Then, the weights and biases of the semantic neural network are optimized using the stochastic gradient descent method. After repeated training, Γ is gradually reduced. CE The loss value is set to enhance the network's anti-distortion ability and obtain a trained neural network.

[0054] Furthermore, based on the trained neural network and the optimized IRS reflection phase shift coefficient matrix, the feature priority list index is used to segment and transmit text semantic information: high-priority data streams are transmitted using IRS, and low-priority data streams are transmitted using fading channels, so as to achieve protection of key semantic features and transmission of semantic information.

[0055] The beneficial effects of the present invention are:

[0056] (1) The present invention extracts features from semantic information based on semantic importance and semantic robustness in conjunction with contextual semantics, and performs weighted combination measurement of semantic features based on dual feature dimensions to achieve effective identification of key semantic features.

[0057] (2) The present invention uses an alternating optimization algorithm to solve the optimal transmit beamforming precoding vector and IRS reflection phase shift coefficient matrix, thereby improving the channel environment and achieving effective protection of key semantic features with high importance and low robustness;

[0058] (3) The present invention proposes a cross-layer optimization method, which performs a joint cross-layer optimization design from the semantic layer and the physical layer, maximizes the received signal power and minimizes the cross-entropy loss of the network function, maximizes the matching degree between the input text and the estimated text, and thus improves the robustness and semantic similarity of the semantic communication system.

[0059] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0061] Figure 1 This is the semantic feature extraction model diagram of the semantic level proposed by the present invention;

[0062] Figure 2 Flowchart for implementing the cross-layer mapping mechanism between semantic features and physical channels proposed in the present invention;

[0063] Figure 3 This is a flowchart of the implementation process of the semantic communication cross-layer design method based on intelligent reflective surfaces proposed in the present invention. DETAILED DESCRIPTION

[0064] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0065] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0066] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0067] See also Figures 1 to 3 , a cross-layer design method for semantic communication based on smart reflective surfaces.

[0068] like Figure 1 As shown, the present invention uses the semantic association loss caused by words in a sentence to measure semantic importance, and uses the incremental loss of perturbed word embedding vectors through a semantic communication system to measure semantic robustness. Before end-to-end transmission, the semantic importance score list and the semantic robustness score list of the transmitted text are respectively fed into the background knowledge base to update the prior semantic knowledge. The sum is then weighted and combined to define a feature priority list containing index information, thereby guiding the channel transmission process of data streams of different priorities.

[0069] like Figure 2 As shown in the figure, in order to reduce semantic-level loss and further reduce the impact of physical noise on semantic features, thereby transmitting semantic information more efficiently, the present invention proposes an optimization method for semantic importance and robustness and combines IRS for cross-layer design. Based on the text source, an end-to-end semantic communication system is designed. The semantic communication system mainly includes four parts: background knowledge base, transmitter, channel, and receiver. The background knowledge base contains various semantic features, which will guide the semantic feature extraction and feature arrangement process. The transmitter includes a semantic encoder for extracting semantic features and a channel encoder for combating channel noise, as well as a feature arrangement module for segmenting the encoded data according to the background knowledge base; the receiver consists of a feature arrangement inverse module for data stream reorganization, a channel decoder for recovering the original data, and a semantic decoder for text estimation.

[0070] like Figure 3As shown in the figure, the system design specifically includes the following steps:

[0071] S1. Use the BERT pre-trained model and combine it with cosine similarity to evaluate the semantic relevance loss of each word to measure the semantic importance of the transmitted text;

[0072] S2. Calculate the mean square error of the loss increment generated by the semantic communication system before and after the feature vector is perturbed, thereby measuring the semantic robustness of the transmitted text;

[0073] S3. Constructing a joint measurement model based on the weighted combination of two-dimensional features to determine the feature priority of the transmitted text and divide it into high-priority data stream and low-priority data stream;

[0074] S4. Introduce a joint function of semantic matching between the physical layer and the semantic layer, and decouple it into a channel function that maximizes signal power and a network function that minimizes cross entropy loss;

[0075] S5. Use an alternating optimization algorithm to maximize the received signal power and obtain the optimal dynamic beamforming precoding vector and IRS passive phase shift coefficient matrix;

[0076] S6. Use stochastic gradient descent to minimize cross entropy loss, repeatedly train and optimize the weights and biases of the semantic neural network, and enhance the network's ability to resist distortion;

[0077] S7. Based on the maximization of semantic matching, high-priority data streams are transmitted using IRS, and low-priority data streams are transmitted using fading channels, thus achieving efficient transmission of semantic information.

[0078] The present invention describes the above method in detail according to three aspects: channel transmission model design, semantic feature extraction, and cross-layer joint optimization.

[0079] (1) Channel transmission model design

[0080] The transmitter first cleans and preprocesses the text data set, and then sends the text s as input. After the input text passes through the semantic encoder and channel encoder, the output data stream x can be expressed as

[0081]

[0082] in, represents the semantic encoder with the network parameter set α; represents the channel encoder with the network parameter set β. The data stream x is prioritized by the feature permutation module, i.e.

[0083] x h ,x l =F T (x; σ)

[0084] Where σ represents the feature priority score list, F T (·) represents the feature ranking module that uses σ to perform priority division at the sending end.

[0085] The channel is designed as two fading channels according to the priority of the data stream: IRS assisted and non-IRS assisted. l When transmitting through a fading channel without IRS assistance, the received low priority data stream y l for

[0086] y l =h d x l +n

[0087] in, represents the Rayleigh fading channel, Variance Complex additive Gaussian white noise.

[0088] To reduce channel fading for high priority data stream x h The impact of recovery, ensuring that higher priority semantic features can be accurately restored, high priority data stream x h Use IRS to assist in fading channels. Assume that the transmitting antenna has N array elements and the IRS has M passive reflective elements. Each reflective element includes an element that can adjust the phase and amplitude of the corresponding incident wave. The direct link from the transmitter to the receiver is represented by H dr ∈C N×1 The equivalent baseband channels from the transmitter to the IRS and from the IRS to the receiver are represented as H T ∈C M×N and H R ∈C M×1 The reflection element set of IRS is represented by M, and the reflection coefficient matrix of IRS can be expressed as

[0089]

[0090] Among them, for each element β m ∈[0,1] and θ m ∈[0,2π) represent the amplitude and phase shift of the mth element of IRS, respectively.

[0091] Assuming that the channel state information is completely known at the transmitter through ideal channel estimation, if the direct channel power is higher than the IRS cascade channel power, the system will prioritize the direct channel, otherwise it will select the IRS cascade channel. Generally, beamforming performs linear precoding, x h The complex baseband transmission signal after beamforming at the transmitting end is expressed as x h =wt c, where c is the data to be transmitted, w t ∈C N×1 represents the beamforming precoding matrix, subject to the maximum available transmit power constraint ||w t || 2 ≤P max After the direct link from the transmitter to the receiver and the IRS cascade link transmission, the high priority data stream y h It can be expressed as

[0092]

[0093] in, The variance is Complex additive Gaussian white noise.

[0094] The transmitted data stream reaches the receiving end after passing through two fading channels with and without IRS assistance. The original permutation signal is restored according to the index through the feature permutation inverse module.

[0095] y=F R (y h ,y l ; σ)

[0096] Among them, F R (·) represents the feature permutation inverse module that uses σ to recover the signal at the receiving end, and y represents the recovered signal at the receiving end.

[0097] The original semantics are restored by the channel decoder and semantic decoder, which estimates the text It can be expressed as

[0098]

[0099] in, and They are the channel decoder and semantic decoder with network parameter sets δ and χ respectively.

[0100] (2) Semantic feature extraction

[0101] The important semantics in the semantic communication system are of great significance to the understanding of the overall semantics. The semantic association loss based on cosine similarity measures the importance of words in the text source. First, the cleaned and preprocessed text s is fed into the pre-trained language model BERT to extract the vector representation B of each word based on the BERT model. ψ (·). These vectors capture rich semantic information, including but not limited to lexical meaning, grammatical role, and contextual relationship. Then, based on the extracted word vector representation B ψ (·) Use cosine similarity to calculate the semantic relevance between two words in the lth sentence in s

[0102]

[0103] Among them, ψ(s l,i ,s l,j ) is the semantic relevance between the i-th and j-th words in the l-th sentence. The superscript T represents the matrix transpose and ||·|| represents the L1 norm. Semantic relevance ψ(s l,i ,s l,j )∈(0,1), semantic relevance ψ(s l,i ,s l,j ) is closer to 1, indicating that the two words are more closely related.

[0104] To accurately quantify the semantic contribution of a word in the context, we need to calculate the sum of the semantic associations between the target word and the remaining words one by one and normalize them, thereby establishing a mapping relationship between the target word and all the words in the sentence. Therefore, the average semantic association between the target word and all the words in the sentence is expressed as follows:

[0105]

[0106] Among them, d represents the length of the sentence, and the average semantic relevance Ψ(s l,i )∈(0,1).

[0107] Based on the above semantic relevance calculation method, the semantic relevance loss caused by a word in a sentence is used to define its semantic importance: when a word has a significant semantic relevance to the context words, removing the word may cause a large semantic change, indicating that the word has a high semantic contribution to the semantic structure of the sentence. For the i-th word in the l-th sentence in s, its semantic importance score can be measured as follows

[0108] I(s l,i )=1-Ψ(s l,i )

[0109] After traversing the input text dataset of length c, the semantic importance score of each word is sorted in descending order and recorded as I = [I1, I2, ..., I c ], and then send I into the background knowledge base for updating.

[0110] The ability of semantic information to tolerate noise in a semantic communication system is closely related to semantic robustness. Higher semantic robustness means that semantic information has a stronger ability to suppress distortion in the presence of channel noise. Semantic robustness is measured by calculating the mean square error (MSE) of the incremental loss incurred by the semantic communication system before and after perturbation of the feature vector.

[0111] The preprocessed text s is passed through the semantic encoder After being mapped into a high-dimensional feature vector, the i-th feature dimension represented by the k-th word high-dimensional feature vector is recorded as

[0112] right After applying additive perturbation, the i-th feature dimension of the high-dimensional perturbation feature vector obtained after the channel codec is It can be expressed as

[0113]

[0114] in, represents the additive Gaussian perturbation noise applied to the high-dimensional feature vector of the k-th word, Denotes the network parameter set as a delta channel decoder, Denotes the channel encoder with the network parameter set β.

[0115] By analyzing the changes in high-dimensional feature vectors caused by perturbations, we establish a mapping of the word's ability to suppress channel noise distortion: the incremental loss of each feature dimension directly reflects the noise tolerance of different semantic units, thereby characterizing their semantic robustness level.

[0116] The loss increment is measured by MSE

[0117]

[0118] Where n represents the number of feature dimensions after the high-dimensional feature vector is mapped by the semantic encoder, Γ k represents the loss increment of the k-th word.

[0119] If the loss increment Γ of the feature vector corresponding to the word k The larger the value is, the worse its semantic robustness is. Then the semantic robustness score r of the kth word is k Defined as Γ k The reciprocal of

[0120]

[0121] After traversing the input text dataset of length c, the semantic robustness scores of each word are sorted in ascending order and recorded as r = [r1, r2, ..., r c ] and sends r to the background knowledge base for updating. It should be noted that the MSE function used in calculating the semantic robustness score does not involve the calculation of neural network gradients or symbol-level reconstruction error; it only measures the incremental loss of semantic information. Furthermore, since incremental loss can only be calculated after transmission, and semantic robustness, as prior knowledge, must be calculated before transmission, this paper uses a pretrained baseline model to measure semantic robustness before end-to-end training.

[0122] After calculating the semantic importance score and semantic robustness score, a joint measurement model needs to be constructed to determine the transmission priority driven by semantic features. Data streams with higher transmission priority are transmitted in fading channels with the assistance of IRS, which can effectively suppress the distortion effect of channel fading on features with high semantic importance or low semantic robustness. This paper measures the transmission priority based on a weighted combination of two-dimensional features. The feature priority score of the k-th word is expressed as

[0123] σ k =λ·I k +μ·(1-r k )

[0124] Among them, I k and r k Represent the semantic importance score and semantic robustness score of the kth word, respectively, λ and μ represent the system's preference coefficients for semantic importance and semantic robustness, respectively. In order to make the formula meaningful, I k and r k They have all been normalized in the interval [0,1], and the coefficients λ and μ are both positive and satisfy the sum of 1. Finally, the feature priority scores of each word are arranged in descending order and recorded as σ=[σ1,σ2,…,σ c ] and send σ into the background knowledge base for updating.

[0125] (3) Cross-layer joint design

[0126] In order to reduce the impact of channel fading on the recovery of high-priority data streams, it is necessary to jointly design the physical layer channel and the semantic layer network across layers. At the physical layer, in order to transmit high-priority data streams as completely as possible, it is necessary to maximize the received signal power. According to the channel model assumed in (1), the signal power received at the receiving end can be expressed as

[0127]

[0128] At the semantic layer, in order to restore the semantics as accurately as possible, the cross entropy loss function is selected as the cost function to minimize the difference between the input text s and the estimated text The difference between them can be expressed as

[0129]

[0130] Among them, q(w l ) is the lth word w in the input text s l The true probability of occurrence, p(w l ) is the estimated text The first word w in l The predicted probability of occurrence, β, α, χ, δ are all neural network hyperparameters.

[0131] To characterize the matching degree between the estimated text and the input text after the system recovers, a joint function combining the physical layer and the semantic layer is introduced, which is named the semantic matching function Mat(·). Mat(·) maximizes the matching degree between the input text and the estimated text by maximizing the received signal power and minimizing the cross entropy loss, which can be expressed as

[0132]

[0133] stC1:||w t || 2 ≤P max ,

[0134]

[0135] C3:β,α,χ,δ are all hyperparameters.

[0136] Solving P0 for the joint function Mat(·) requires a joint cross-layer design of the network parameters in the semantic layer and the beamforming vectors and IRS phase shifts in the physical layer channel. It is not difficult to find that the cross-layer optimization variables have a natural decoupling characteristic: the physical layer optimization variables focus on maximizing signal strength under certain channel conditions; the semantic layer optimization variables are dedicated to adjusting the semantic codec parameters of the neural network to minimize semantic loss. In addition, C1 limits the transmit power, C2 constrains the IRS phase shift unit modulus, and C3 limits the network hyperparameters. These three constraints act on different parameter spaces and have orthogonal solution spaces. Therefore, Mat(·) can be separated and solved separately.

[0137] The corresponding sub-problems in the physical layer of the joint function can be separated to obtain Mat(γ R ,Γ CE )=γ R The problem is formulated as follows: To enhance wireless transmission over the physical channel, it is necessary to jointly design the active beamforming precoding vector w t and the passive phase shift coefficient θ of the IRS n , to maximize the received signal power γ R

[0138]

[0139] stC1:||w t || 2 ≤P max ,

[0140]

[0141] For the above problem, an alternating optimization algorithm is used to solve it: the transmit beamforming vector at the transmitter and the phase shift at the IRS are iteratively optimized in an alternating manner, with one fixed at each iteration until the two converge or the maximum number of iterations is reached, and the optimal transmit beamforming precoding vector is obtained. and the reflection phase shift coefficient matrix Φ * .

[0142] It is not difficult to find that the optimal solution to this problem using alternating optimization is to tune the phase shift of the nth element so that the signal phase of the IRS cascade link is aligned with the signal phase of the direct link, thereby achieving coherent signal combination at the user. Therefore, the optimal IRS reflection phase shift coefficient matrix Φ * The phase shift of the nth element in is given by

[0143]

[0144] The receiving end signal is represented by the direct link and the IRS cascade link as Therefore, the optimal transmit beamforming solution is It can be given by Maximum Ratio Transmission (MRT):

[0145]

[0146] The above alternating optimization method does not require calling a semidefinite programming solver. The transmit beamforming vector and IRS phase shift are both obtained in closed-form expressions and can therefore be directly applied to the system for end-to-end training.

[0147] After solving the problem of maximizing the received signal power at the physical layer, it is necessary to train the neural network parameters of the semantic layer. The sub-problem of training the cross entropy loss function under the C3 constraint of the semantic layer is separated from the joint function and expressed as:

[0148]

[0149] stC3:β,α,χ,δ are all hyperparameters

[0150] This subproblem aims to minimize the semantic information error and uses the cross entropy function as the loss function to quantify the difference between the input text sequence and the output text sequence. First, the loss at the receiving end is calculated and back-propagated to the sending end. Then, the weights and biases of the semantic neural network are optimized using the stochastic gradient descent method. After repeated training, Γ is gradually reduced. CE The loss value of , makes the network's anti-distortion ability gradually enhanced, and finally a well-trained neural network is obtained.

[0151] After solving the two problems of maximizing the received signal power at the physical layer and minimizing the cross-entropy loss at the semantic layer, the system was trained and tested on a large scale using text datasets, and the relevant parameters were continuously adjusted. Finally, the optimal cross-layer jointly designed semantic communication system based on intelligent reflective surfaces was obtained.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A cross-layer design method for semantic communication based on intelligent reflective surfaces, characterized by: The following steps are involved: S1: Use a bidirectional encoder to represent the BERT pre-trained model and combine it with cosine similarity to evaluate the semantic relevance loss of each word to measure the semantic importance of the transmitted text; S2: Based on the mean square error (MSE) of the incremental loss generated by the semantic communication system before and after the feature vector perturbation, a mapping of the word's ability to suppress channel noise distortion is established, thereby characterizing the semantic robustness of the transmitted text; S3: Based on the weighted combination of two-dimensional features, a joint measurement model is constructed to determine the feature priority of the transmitted text and divide it into high-priority data streams and low-priority data streams; S4: A joint function of the physical layer and the semantic layer is introduced to characterize the matching degree between the recovered estimated text and the input text, and it is decoupled into a channel function that maximizes signal power and a network function that minimizes cross entropy loss; S5: Use an alternating optimization algorithm to maximize the received signal power and obtain the optimal dynamic beamforming precoding vector and IRS passive phase shift coefficient matrix; S6: With the goal of minimizing semantic information errors, the stochastic gradient descent method is used to minimize cross-entropy loss. The weights and biases of the semantic neural network are repeatedly trained to enhance the network's anti-distortion capabilities. Based on the optimized parameters, the text semantic information is transmitted. S7: Transmit text semantic information based on the feature priority list index and the optimized IRS reflection phase shift coefficient matrix: Use IRS to transmit high-priority data streams and fading channels to transmit low-priority data streams, thereby achieving favorable protection of key semantic features and transmission of semantic information.

2. The cross-layer semantic communication design method based on intelligent reflective surfaces according to claim 1, characterized in that: Said S1 comprises the following steps: The preprocessed text s is fed into the pre-trained language model BERT, and the vector representation B of each word based on the BERT model is extracted. ψ (·), then based on the extracted vector representation B ψ (·) Use cosine similarity to calculate the semantic relevance between two words in the lth sentence: Where i and j represent the i-th and j-th words in the sentence respectively, the superscript T represents the matrix transpose, ||·|| represents the L1 norm, ψ(s l,i ,s l,j ) represents the semantic relevance between the i-th and j-th words in the l-th sentence; In order to accurately quantify the semantic contribution of a word in the context, the average semantic association between the target word and all other words in the sentence is established as follows: Among them, d represents the length of the sentence, Ψ(s l,i ) represents the average semantic relevance of the i-th word in the l-th sentence; Based on the above semantic relevance calculation method, the semantic relevance loss caused by a word in a sentence is used to define its semantic importance: when a word has a significant semantic relevance to the context words, removing the word will lead to a large semantic change, which indicates that the word has a high semantic contribution in the semantic composition of the sentence. For the i-th word in the l-th sentence in s, its semantic importance score I(s) is l,i ) is expressed as: I(s l,i )=1-Ψ(s l,i )。 3. The cross-layer semantic communication design method based on intelligent reflective surfaces according to claim 2 is characterized by: The S2 comprises the following steps: The preprocessed text s is passed through the semantic encoder Then it is mapped into a high-dimensional feature vector right After applying additive perturbation, the i-th feature dimension of the perturbation feature vector obtained after the channel codec is expressed as The loss increment of each feature dimension directly reflects the noise tolerance of different semantic units, and thus characterizes their semantic robustness level. The loss increment is measured by MSE: Among them, n represents the feature dimension size mapped after the semantic encoder, Γ k represents the loss increment of the k-th word; If the loss increment Γ of the feature vector corresponding to the word k The larger the value is, the worse its semantic robustness is. Then the semantic robustness score r of the kth word is k Defined as Γ k The reciprocal of:

4. The cross-layer semantic communication design method based on intelligent reflective surfaces according to claim 3, characterized in that: The S3 includes the following steps: s k =λ·I k +μ·(1-r k ) Among them, I k and r k denote the semantic importance score and semantic robustness score of the kth word, λ and μ denote the system's preference coefficients for semantic importance and semantic robustness, σ k represents the feature priority score of the kth word in the text; Finally, the feature priority scores of each word are arranged in descending order and recorded as σ=[σ1,σ2,…,σ c ], and send σ into the background knowledge base. During transmission, the text data stream will be divided into high-priority data stream and low-priority data stream according to the feature priority score list.

5. The cross-layer design method for semantic communication based on intelligent reflective surfaces according to claim 1, characterized in that: In S4, in order to characterize the matching degree between the estimated text after the system recovery and the input text, a joint function combining the physical layer and the semantic layer is introduced, which is named the semantic matching function Mat(·): s.t.C1:||w t || 2 ≤P max , C3:β,α,χ,δ are all hyperparameters Among them, w t represents the beamforming precoding vector, Φ represents the IRS reflection phase shift coefficient matrix, γ R represents the signal power, Γ CE represents the cross entropy loss, P max represents the maximum power constraint, n represents the nth IRS element, θ n represents the phase shift of the nth element in the IRS reflection phase shift coefficient matrix, C1 represents constraint 1, C2 represents constraint 2, and C3 represents constraint 3; By maximizing the received signal power and minimizing the cross-entropy loss, the matching degree between the input text and the estimated text is maximized, and then it is decoupled into a channel function that maximizes the signal power and a network function that minimizes the cross-entropy loss.

6. The cross-layer design method for semantic communication based on intelligent reflective surfaces according to claim 5, characterized in that: The corresponding sub-problem in the physical layer of the joint function is decoupled into the channel function that maximizes the signal power and solved separately. It can be expressed as: to enhance wireless transmission on the physical channel, it is necessary to jointly design the active beamforming precoding vector w t and the passive phase shift coefficient θ of the IRS n , to maximize the received signal power γ R : s.t.C1:||w t || 2 ≤P max , Among them, |·| represents the absolute value, H dr ∈C N×1 Indicates the direct link from the transmitter to the receiver, H T ∈C M×N and H R ∈C M×1 Denote the equivalent baseband channels from the transmitter to the IRS and from the IRS to the receiver, respectively. The set of reflection elements of the IRS is represented by M. For each element β m ∈[0,1] and θ m ∈[0,2π) represent the amplitude and phase shift of the mth element of IRS respectively. The reflection coefficient matrix of IRS is expressed as For the above problem, an alternating optimization algorithm is used to solve it. The transmit beamforming vector at the transmitter and the phase shift at the IRS are iteratively optimized in an alternating manner. One is fixed each time until the two converge or the maximum number of iterations is reached, and the optimal transmit beamforming precoding vector is obtained. and the reflection phase shift coefficient matrix Φ * .

7. The cross-layer semantic communication design method based on intelligent reflective surfaces according to claim 5, characterized in that: The corresponding sub-problems in the semantic layer of the joint function are decoupled into network functions that minimize the cross entropy loss and solved separately: stC3:β,α,χ,δ are all hyperparameters This subproblem aims to minimize the semantic information error and uses the cross entropy function as the loss function to quantify the difference between the input text sequence and the output text sequence; First, the loss of the receiving end is calculated and back-propagated to the sending end. Then, the weights and biases of the semantic neural network are optimized using the stochastic gradient descent method. After repeated training, Γ is gradually reduced. CE The loss value is set to enhance the network's anti-distortion ability and obtain a trained neural network.

8. The cross-layer semantic communication design method based on intelligent reflective surfaces according to claim 7, characterized in that: Based on the trained neural network and the optimized IRS reflection phase shift coefficient matrix, the feature priority list index is used to segment and transmit text semantic information: high-priority data streams are transmitted using IRS, and low-priority data streams are transmitted using fading channels, so as to achieve the protection of key semantic features and the transmission of semantic information.

Images