An end-to-end semantic reconstruction method and device for a low signal-to-noise ratio scene

Abstract

The application discloses a kind of end-to-end semantic reconstruction methods and devices for low signal-to-noise ratio scene, belong to communication security research field technical field.The original message semantic feature is extracted by sending end, and symbol stream is generated by semantic coding and channel coding and sent;Channel decoding and semantic decoding are carried out by receiving end, and damaged feature and signal-to-noise ratio estimation are obtained;Original message and signal-to-noise ratio estimation are input into channel attention encoder, and target semantic representation is generated;Condition diffusion model is constructed, and damaged feature is used as constraint, and reconstruction semantic representation aligned with target semantic representation is generated in reverse diffusion;It is input into channel attention decoder with signal-to-noise ratio estimation, and the message of adaptive decoding recovery is recovered.The application realizes semantic high-fidelity reconstruction under low signal-to-noise ratio, improves semantic transmission accuracy and system robustness.

Description

Technical Field

[0001] This invention relates to the technical field of communication security research, and in particular to an end-to-end semantic reconstruction method, apparatus, device, and storage medium for low signal-to-noise ratio scenarios. Background Technology

[0002] Semantic communication, as an emerging communication paradigm integrating cutting-edge technologies such as natural language processing, machine learning, and artificial intelligence, focuses on the semantic connotations carried by the transmitted information rather than the traditional bit-level symbols themselves. By extracting and modeling semantic features at the transmission level, semantic communication can achieve a certain degree of semantic fidelity at the receiving end, demonstrating higher transmission efficiency and stronger task adaptability compared to traditional communication systems in complex and ever-changing application scenarios. In recent years, with the in-depth application of deep neural networks and end-to-end learning mechanisms in communication systems, semantic communication has achieved rapid development in fields such as intelligent transportation, telemedicine, and the Industrial Internet of Things, becoming a research hotspot of common interest to both academia and industry.

[0003] However, although semantic communication theoretically allows for a certain degree of bit-level transmission error and can partially complete information through high-level semantic association, in real-world wireless environments, especially under low signal-to-noise ratio (SNR) conditions, channel noise has a cumulative and asymmetric effect on semantic features. The deep semantic features obtained by the receiver through channel decoding and semantic decoding often suffer from severe local missing features, structural distortions, or distribution shifts. Directly using these damaged features for downstream semantic task inference not only makes it difficult to effectively recover the original semantic information but also leads to semantic comprehension biases, task execution failures, and even introduces systemic security vulnerabilities. Against this backdrop, how to reconstruct damaged semantic features with high quality in low SNR scenarios has become one of the core bottlenecks restricting the engineering implementation of semantic communication.

[0004] Existing semantic information reconstruction techniques can be broadly categorized into three types: The first type is based on traditional signal processing methods, improving received signal quality through channel estimation, equalization, and error correction coding. However, these methods lack structural priors at the semantic level, making it difficult to achieve information completion from a semantic distribution perspective. The second type is end-to-end reconstruction methods based on deep semantic coding, achieving semantic feature recovery through joint optimization of the encoding and decoding network. However, these methods typically assume known and stable channel states, lacking adaptability to dynamic channel environments, and their reconstruction performance significantly degrades when the signal-to-noise ratio fluctuates drastically. The third type introduces generative models such as generative adversarial networks or variational autoencoders, achieving feature reconstruction by modeling the semantic distribution. However, existing methods are prone to pattern collapse or insufficient alignment between the generated results and the true semantic distribution under low signal-to-noise ratio conditions, and their computational complexity is high, making them difficult to meet the needs of real-time communication.

[0005] In summary, existing technologies still face multiple challenges in low signal-to-noise ratio scenarios, such as severe damage to semantic features, insufficient channel state adaptation capability, and limited semantic fidelity. There is an urgent need to propose an end-to-end semantic reconstruction mechanism that can integrate channel state awareness and semantic distribution constraints to achieve high-fidelity recovery of semantic information under low signal-to-noise ratio conditions. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] To address this, this invention proposes an end-to-end semantic reconstruction method for low signal-to-noise ratio (SNR) scenarios. The sending end extracts the semantic features of the original message and generates a symbol stream through semantic coding and channel coding before transmission. The receiving end performs channel decoding and semantic decoding to obtain the damage depth features and SNR estimation. The original message and SNR estimation are input into a channel attention encoder to generate a target semantic representation modulated by channel state. A conditional diffusion model is constructed, using the damage depth features as constraints, to progressively denoise during back-diffusion, generating a reconstructed semantic representation aligned with the target semantic representation. The reconstructed semantic representation and SNR estimation are input into a channel attention decoder to perform channel state adaptation decoding, recovering the final message. This invention achieves high-fidelity semantic information reconstruction under low SNR conditions, significantly improving the accuracy of semantic transmission and system robustness.

[0008] Another objective of this invention is to provide an end-to-end semantic reconstruction device for low signal-to-noise ratio scenarios.

[0009] The third objective of this invention is to provide a computer device.

[0010] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0011] To achieve the above objectives, this invention proposes an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios, comprising:

[0012] The sending end extracts semantic features from the original message to generate deep semantic features to be transmitted, and performs semantic coding and channel coding on the deep semantic features in sequence to generate a symbol stream for transmission through the wireless channel. By performing channel decoding and semantic decoding sequentially on the received signal at the receiving end, the damage depth characteristics are obtained, and the current signal-to-noise ratio estimate is obtained through channel estimation. The original message and the current signal-to-noise ratio estimate are input together into the channel attention encoder to generate a target semantic representation modulated by the channel state. A conditional diffusion model is constructed. During the back diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples and generate a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution. The reconstructed semantic representation and the current signal-to-noise ratio estimate are input into the channel attention decoder. The reconstructed semantic representation is then subjected to a decoding operation adapted to the channel state to recover the final message.

[0013] An end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, the step of extracting semantic features from the original message at the sending end to generate deep semantic features to be transmitted, and sequentially performing semantic coding and channel coding on the deep semantic features to generate a symbol stream for transmission through a wireless channel includes: The original message is input into a pre-trained feature extraction network at the sending end. The deep semantic features of the original message in the latent space are extracted through multi-layer nonlinear transformation. The deep semantic features are then input into a pre-trained semantic encoder. The semantic encoder uses a multi-head attention mechanism to model the global dependency relationship and compress the dimension of the deep semantic features, and outputs a fixed-length semantic feature vector. The semantic feature vector is input to a pre-trained channel encoder, which maps the semantic feature vector into a real-domain symbol stream suitable for physical channel transmission, applies power normalization constraints to the real-domain symbol stream, and then transmits the real-domain symbol stream to the receiver via a wireless channel.

[0014] In one embodiment of the present invention, the step of sequentially performing channel decoding and semantic decoding on the received signal at the receiving end to obtain the impairment depth features, and obtaining the current signal-to-noise ratio estimate through channel estimation, includes: The receiver performs synchronization and channel equalization on the received symbol stream, and inputs the recovered baseband symbol sequence into a pre-trained channel decoder. The channel decoder inversely maps the baseband symbol sequence into an initial estimation vector in the semantic feature space, and inputs the initial estimation vector into a pre-trained semantic decoder. The semantic decoder adopts an attention decoding structure that is symmetrical to the semantic encoder at the transmitter, upsamples the initial estimation vector and reconstructs semantic relations, and outputs a damaged depth feature in the latent space that is consistent with the dimension of the deep semantic feature at the transmitter. By using the pilot symbols and reference signals embedded in the received signal at the receiving end, the channel quality is measured. The instantaneous signal-to-noise ratio of the current channel is estimated by the least squares and least mean square error criteria, and the current signal-to-noise ratio estimate is generated. Then, the damage depth feature and the current signal-to-noise ratio estimate are used as the common input of the subsequent conditional diffusion model and channel attention decoder, so that the receiving end reconstruction process depends on both the semantic content of the signal itself and the statistical characteristics of the channel state.

[0015] In one embodiment of the present invention, the step of inputting the original message and the current signal-to-noise ratio estimate together into the channel attention encoder to generate a channel state modulated target semantic representation includes: The receiver performs numerical normalization on the current signal-to-noise ratio (SNR) estimate obtained through channel estimation, mapping the current SNR estimate to a feature space dimension that is identical and alignable to the semantic embedding representation of the original message, thus generating an SNR embedding vector. Then, the semantic embedding representation of the original message and the SNR embedding vector are concatenated along the feature dimension and fused element-wise with weights to form a joint representation containing semantic content and channel state information. This joint representation is input to the channel attention encoder, which includes a multi-head cross-attention layer. Using the joint representation as the query and the semantic embedding representation as the key and value, the encoder calculates the SNR-conditionally guided attention weight distribution. Adaptive weighted modulation is applied to the semantic embedding representation of the original message according to the attention weight distribution, so that the components of the semantic features that are sensitive to channel noise are suppressed and enhanced to different degrees. The target semantic representation is output after explicit calibration of the channel state. The target semantic representation is used as the initial sample in the forward diffusion process of the subsequent conditional diffusion model during the training phase, and as the ideal reconstruction target to be approximated in the backward diffusion process.

[0016] In one embodiment of the present invention, the construction of a conditional diffusion model, wherein the conditional diffusion model, during the back-diffusion process, uses the damaged depth feature as a conditional constraint to perform progressive denoising on the noisy samples, generating a reconstructed semantic representation aligned with the target semantic representation in terms of semantic distribution, includes: A predefined forward diffusion process is used, with the target semantic representation output by the channel attention encoder as the initial sample. Gaussian noise is iteratively added to the initial sample according to a predefined variance scheduling table, generating a Markov chain from clean samples to a pure noise distribution; the noisy sample at step t is represented as: ; in, For the target semantic representation, For cumulative noise dispatch coefficient, For the noisy sample added at step t; A conditional denoising network is constructed. The conditional denoising network takes the noisy sample of the current step, the index of the current time step, and the damaged depth features as input. Through cross-layer connections and attention modulation, the damaged depth features are embedded into each decoding layer of the network, so that the denoising process is explicitly guided by the semantic content actually recovered by the receiver. During the backdiffusion process, starting from random noise and the maximum noise step sample obtained through forward diffusion, the fixed-condition denoising network is called to iteratively predict the noise component added in the current step, using the damaged depth feature as a fixed condition. The noise is gradually removed and the semantic structure is restored, generating an intermediate state semantic representation sequence. The semantic representation output at the backdiffusion termination step is used as the reconstructed semantic representation. The reconstructed semantic representation satisfies the preset similarity threshold with the target semantic representation corresponding to the original message at the sending end in the semantic embedding space, thus achieving probability alignment from damaged semantic features to complete semantic distribution. During the training phase, the diffusion model loss function is constructed with the objective of minimizing the mean square error between the noise predicted by the conditional denoising network and the actual Gaussian noise: ; in, For parameter set Conditional denoising network, The damaged depth characteristics recovered by the receiving end.

[0017] In one embodiment of the present invention, the step of inputting the reconstructed semantic representation and the current signal-to-noise ratio estimate into the channel attention decoder, performing a decoding operation adapted to the channel state on the reconstructed semantic representation, and recovering the final message includes: The receiver normalizes the current signal-to-noise ratio (SNR) estimate and performs a linear embedding transformation to generate an SNR conditional vector aligned with the dimensions of the reconstructed semantic representation. The reconstructed semantic representation and the SNR conditional vector are concatenated along the feature dimension, and a channel-wise affine transformation is used to modulate the SNR conditional vector to the mean and variance of the reconstructed semantic representation, forming a channel-state-aware joint semantic representation. This joint semantic representation is then input to a channel attention decoder, which includes a cross-attention module. Using the joint semantic representation as the query and the pre-stored codebook and the prototype embedding of the transmitter's semantic space as the key and value, the module calculates the SNR-guided attention weights to perform feature calibration and semantic completion on the reconstructed semantic representation. The calibrated reconstructed semantic representation is passed sequentially through the semantic decoding layer and the feature inverse transformation layer, and upsampling, nonlinear mapping and discretization operations symmetrical to the feature extraction network at the sending end are performed to recover the final message sequence consistent with the original message format. During the joint training phase, the channel attention decoder, conditional diffusion model, and channel attention encoder undergo collaborative parameter optimization using an end-to-end loss function. This end-to-end loss function employs the cross-entropy criterion to measure the semantic deviation between the final message and the original message. ; in, The original message from the sender. The final message recovered by the receiving end. The parameter set for the feature extractor For the parameter set of the semantic encoder, For the parameter set of the channel encoder, For the parameter set of the channel decoder, For the parameter set of the semantic decoder, For the parameter set of the channel attention encoder, For the parameter set of the channel attention decoder, This is the parameter set for the conditional diffusion model; Through the dynamic adjustment mechanism of the attention weights inside the channel attention decoder, the decoding process adaptively switches the degree of dependence on the reconstructed semantic representation under different signal-to-noise ratio (SNR) conditions. When the SNR estimate is higher than the threshold, the decoder focuses on directly utilizing the detailed features in the reconstructed semantic representation; when the SNR estimate is lower than the threshold, the decoder enhances its dependence on the semantic prototype embedding, suppresses feature distortion introduced by noise, and achieves robust semantic recovery with channel state awareness.

[0018] To achieve the above objectives, another aspect of the present invention proposes an end-to-end semantic reconstruction apparatus for low signal-to-noise ratio scenarios, comprising: The sending end semantic encoding and transmission module is used to extract semantic features from the original message through the sending end, generate deep semantic features to be transmitted, and sequentially perform semantic encoding and channel encoding on the deep semantic features to generate a symbol stream for transmission through the wireless channel. The receiver initial decoding and channel estimation module is used to perform channel decoding and semantic decoding on the received signal sequentially through the receiver to obtain the damage depth features, and to obtain the current signal-to-noise ratio estimate through channel estimation. The channel state modulation module is used to input the original message and the current signal-to-noise ratio estimate into the channel attention encoder to generate a target semantic representation modulated by the channel state. The conditional diffusion reconstruction module is used to construct a conditional diffusion model. During the back diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples and generate a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution. The channel attention decoding and message recovery module is used to input the reconstructed semantic representation and the current signal-to-noise ratio estimate into the channel attention decoder, perform a decoding operation adapted to the channel state on the reconstructed semantic representation, and recover the final message.

[0019] This invention discloses an end-to-end semantic reconstruction method and apparatus for low signal-to-noise ratio (SNR) scenarios. The method involves: a transmitting end extracting semantic features and performing semantic and channel coding to generate a symbol stream; a receiving end decoding to obtain the depth features of the damaged semantic features and an estimated SNR, inputting the original message and SNR into a channel attention encoder to generate a target semantic representation modulated by channel state; constructing a conditional diffusion model constrained by the damaged features, and progressively denoising to generate a reconstructed semantic representation aligned with the target semantic representation during back-diffusion; and inputting the reconstructed semantic representation and SNR into a channel attention decoder to perform channel state adaptation decoding to recover the final message. This method effectively solves the problems of severely damaged semantic features, low reconstruction fidelity, and insufficient channel state adaptation capability under low SNR conditions, achieving high-fidelity reconstruction and robust recovery of semantic information, and significantly improving the accuracy of semantic transmission and system reliability.

[0020] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios as described in the first aspect embodiment.

[0021] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios as described in the first aspect embodiment.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a reconstruction scheme for another end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios according to an embodiment of the present invention; Figure 3This is a schematic diagram of the structure of an end-to-end semantic reconstruction device for low signal-to-noise ratio scenarios according to an embodiment of the present invention; Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] The following description, with reference to the accompanying drawings, describes an end-to-end semantic reconstruction method, apparatus, device, and storage medium for low signal-to-noise ratio scenarios according to embodiments of the present invention.

[0027] The core idea of ​​this invention is as follows: At the transmitting end, deep semantic features are extracted from the original message, and semantic encoding and channel encoding are performed sequentially to generate a symbol stream for transmission through a wireless channel. At the receiving end, channel decoding and semantic decoding are performed sequentially on the received signal to obtain damaged deep features, and the current signal-to-noise ratio (SNR) estimate is obtained through channel estimation. The original message and the current SNR estimate are input into a channel attention encoder, and the semantic features are adaptively modulated using SNR conditions to generate a target semantic representation explicitly calibrated by the channel state. A conditional diffusion model constrained by damaged deep features is constructed, and noisy samples are progressively denoised during back-diffusion. Using damaged features as a generation guide, the semantic structure is iteratively restored to generate a reconstructed semantic representation aligned with the target semantic representation in semantic distribution. The reconstructed semantic representation and the current SNR estimate are input into a channel attention decoder, and the reconstructed semantic representation is feature-calibrated and semantically completed using a SNR-guided cross-attention mechanism. Decoding operations adapted to the channel state are performed to recover the final message. This invention extends symbol-level recovery in traditional communication to semantic-level reconstruction and alignment. Through the synergy of channel state-aware modulation, diffusion-generative completion, and attention-adaptive decoding, it transforms severely degraded semantic features under low signal-to-noise ratio conditions into high-fidelity reconstructed representations aligned with the original semantic distribution. This effectively solves the problems of severely damaged semantic features, insufficient channel adaptability, and low reconstruction fidelity, achieving robust recovery and accurate transmission of semantic information under harsh channel conditions, and significantly improving the accuracy and reliability of semantic communication systems.

[0028] Example 1 To achieve the above invention, embodiments of the present invention provide an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios, such as... Figure 1 As shown, it includes: S1, the sending end extracts semantic features from the original message to generate deep semantic features to be transmitted, and performs semantic coding and channel coding on the deep semantic features in sequence to generate a symbol stream for transmission through the wireless channel.

[0029] Specifically, at the sending end, semantic features are first extracted from the original message to obtain a deep feature representation that can characterize the core semantic information of the message. This process is based on the fundamental principles of semantic communication and aims to extract high-level semantic abstractions from the source, discarding redundant symbol-level information, thereby improving transmission efficiency and anti-interference capabilities. Specifically, the sending end is equipped with a feature extractor, a semantic encoder, and a channel encoder in sequence. The feature extractor uses a deep neural network structure, such as a Transformer or convolutional neural network, to map the original text or data into a deep semantic feature vector, which contains the key semantic content of the message. Subsequently, the semantic encoder further compresses and encodes this deep feature, capturing the contextual dependencies within the feature through an attention mechanism or recurrent neural network to generate compact semantic encoded features. Finally, the channel encoder converts the semantic encoded features into a symbol stream suitable for transmission over a wireless channel, which is modulated and transmitted through an antenna. The above process can be formally represented as: ; in, Indicates the original message. , , They respectively represent the parameter sets as , , Feature extractor, semantic encoder, and channel encoder, This is the final generated transmission symbol stream.

[0030] Furthermore, the parameters of each of the above-mentioned encoding modules are all learnable neural network weights, optimized through end-to-end joint training. During training, the transmitting end encoding module and the receiving end decoding module update collaboratively, aiming to minimize the difference between the reconstructed message and the original message. This allows the transmitting end to adaptively adjust its encoding strategy and extract semantic features robust to channel noise. Simultaneously, the symbol stream output by the transmitting end must meet power constraints to ensure that the actual transmit power conforms to the communication system specifications.

[0031] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are often harsh, and traditional communication methods struggle to guarantee complete information transmission. However, this invention, through deep extraction and encoding of semantic features at the sending end, can preserve the core semantics of the message at the source, enhancing its anti-interference capability during transmission and providing a solid feature foundation for subsequent high-quality semantic reconstruction based on a conditional diffusion model at the receiving end.

[0032] Specifically, through the aforementioned processing at the transmitting end, the system can maintain the integrity of semantic information in complex channel environments, significantly improving the accuracy and stability of semantic recovery under low signal-to-noise ratio conditions. Compared with existing technologies, this scheme does not rely on simple bit-level protection, but instead performs feature encoding at the semantic level, making the transmitted symbol stream more robust to noise. This effectively overcomes the problem of the sharp performance degradation of traditional communication under harsh channels, laying the foundation for building a highly reliable and efficient semantic communication system.

[0033] Furthermore, S1 includes: S11, the original message is input to the pre-trained feature extraction network through the sending end, and the deep semantic features of the original message in the latent space are extracted through multi-layer nonlinear transformation. The deep semantic features are then input to the pre-trained semantic encoder. The semantic encoder uses a multi-head attention mechanism to model the global dependency relationship and compress the dimension of the deep semantic features, and outputs a fixed-length semantic feature vector.

[0034] Specifically, in the sending end processing flow, the original message is first input into a pre-trained feature extraction network. This network, built on a deep neural network, maps the original message to a latent space through multiple layers of nonlinear transformations to obtain its deep semantic feature representation. The original message typically contains a large amount of redundant symbol-level information, while semantic communication aims to extract high-level semantic abstractions to improve transmission efficiency and robustness to channel noise. The feature extraction network abstracts and reduces the dimensionality of the input data layer by layer through stacked linear transformations and nonlinear activation functions, ultimately generating deep features in the latent space that can represent the core semantics of the message. This feature extraction network needs to be pre-trained on a large-scale corpus or related dataset to ensure it has good semantic representation capabilities, thus providing high-quality input for subsequent semantic encoding and transmission.

[0035] Furthermore, the feature extraction network can employ an encoder structure based on Transformer, or an architecture such as a residual convolutional neural network. Taking Transformer as an example, the original message is first converted into a sequence of word vectors through an embedding layer, and then transformed through a multi-layer self-attention mechanism and a feedforward network. In each layer, the self-attention mechanism captures the dependencies between different positions in the sequence, while the feedforward network independently transforms the representation of each position. After multiple layers are stacked, the network outputs the hidden state vector corresponding to each position, and these hidden state vectors constitute the deep semantic features. This feature not only preserves the semantic information of the original message, but also removes some redundant details through the abstraction capabilities of the deep network, laying the foundation for subsequent semantic encoding.

[0036] Furthermore, the aforementioned deep semantic features are input into a pre-trained semantic encoder. This semantic encoder employs a multi-head attention mechanism to model the global dependencies of the deep semantic features. In its implementation, the semantic encoder takes the hidden state sequence output by the feature extraction network as input and calculates the attention weights between different positions in the sequence in parallel through a multi-head self-attention mechanism, thereby capturing long-distance semantic dependencies. Each attention head focuses on a different representation subspace, enabling the model to aggregate contextual information from multiple perspectives. Subsequently, the attention-weighted features are further transformed through a feedforward network and aggregated through pooling operations or special classification tokens (such as [CLS] tokens), ultimately outputting a fixed-length semantic feature vector. This fixed-length vector is a compact semantic abstraction of the original message, which achieves dimensionality compression while preserving core semantic information, facilitating subsequent channel coding and wireless transmission.

[0037] Specifically, both the feature extraction network and the semantic encoder are parameterized neural network modules. The parameter set of the feature extraction network is denoted as α, and the parameter set of the semantic encoder is denoted as β. Both are optimized on large-scale datasets through a pre-training phase. The pre-training objective is typically a self-supervised task, such as masked language modeling or contrastive learning, enabling the network to learn general semantic representation capabilities. The fixed-length semantic feature vector dimension is a hyperparameter, which needs to be set according to the bandwidth constraints and semantic task complexity in the actual application scenario. This dimension determines the degree of semantic information compression; a lower dimension helps save transmission resources but may lose some detailed information; a higher dimension retains richer semantic details but consumes more channel resources. In subsequent end-to-end joint training, parameters α and β will be fine-tuned in conjunction with other modules to adapt to specific communication tasks and channel conditions.

[0038] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are often harsh, and traditional communication methods struggle to guarantee complete information transmission. Through deep processing of feature extraction and semantic encoding at the sending end, the system can retain the core semantics of the message at the source end and represent it in a compact vector form of fixed length. This enables efficient transmission in environments with limited bandwidth and high noise, providing a solid feature foundation for subsequent semantic reconstruction based on a conditional diffusion model at the receiving end.

[0039] Specifically, through the aforementioned processing at the sending end, the system can maintain the integrity of semantic information in complex channel environments, significantly improving the accuracy and stability of semantic recovery under low signal-to-noise ratio conditions. Compared with existing technologies, this scheme employs a pre-trained feature extraction network and a semantic encoder based on multi-head attention, enabling the sending end to adaptively extract the most critical semantic components from the message and compress them into a fixed-length vector representation. This processing not only reduces the bandwidth consumption of subsequent transmissions but also makes the transmitted semantic features more robust to channel noise, laying a solid foundation for building a highly reliable and efficient semantic communication system.

[0040] S12, the semantic feature vector is input to the pre-trained channel encoder, the channel encoder maps the semantic feature vector into a real number domain symbol stream suitable for physical channel transmission, performs power normalization constraint on the real number domain symbol stream, and then transmits the real number domain symbol stream to the receiving end through the wireless channel.

[0041] Specifically, in the transmitting end processing flow, the fixed-length semantic feature vector generated in the aforementioned steps is input to the pre-trained channel encoder. The function of this channel encoder is to map from the semantic feature space to the physical transmission symbol space, enabling semantic information to be transmitted in a form suitable for wireless channel transmission. Its implementation principle is based on the nonlinear transformation capability of deep neural networks on features. By converting discrete semantic feature vectors into a continuous real-number domain symbol stream, it avoids the independent modulation and coding modules in traditional communication systems, thereby achieving joint source-channel coding at the semantic level and improving end-to-end transmission efficiency.

[0042] Furthermore, the channel encoder can be constructed using a multi-layer fully connected neural network or a convolutional neural network. Taking a fully connected network as an example, the semantic feature vector first passes through several hidden layers, each performing a linear transformation and non-linear activation sequentially, gradually mapping the features to the target dimension. The number of nodes in the network's output layer is equal to the length of the symbol stream to be transmitted, and no activation function is applied; the real-number domain values ​​are directly output, forming the symbol stream to be transmitted. To ensure that the symbol stream meets the actual transmit power limit, power normalization processing is required after the output layer. A typical implementation is to calculate the L2 norm of the symbol vector and scale it to a preset average power value, for example, setting the average power of each symbol vector to 1, i.e.: ; in, This is the original real-valued symbol vector output by the channel encoder. For the normalized transmission symbol stream, This is the set transmit power. This normalization operation ensures that the symbol stream meets the transmitter's power constraints when passing through the wireless channel, avoiding nonlinear distortion or violations of communication regulations due to excessive power.

[0043] Furthermore, the parameter set of the channel encoder is denoted as... The network needs to be optimized during the pre-training phase or end-to-end joint training. During the pre-training phase, the loss function described in the formula can be used to minimize the reconstruction error before and after channel encoding and decoding, enabling the network to learn the basic mapping ability to resist channel noise under different signal-to-noise ratio conditions. The length of the symbol stream (i.e., the output dimension) is an important system parameter and needs to be designed with trade-offs based on available bandwidth, channel conditions, and semantic task requirements. Longer symbol streams consume more time-frequency resources but can carry richer semantic information; shorter symbol streams save resources but may introduce greater compression losses. After power normalization, the transmitter frames the real-number domain symbol stream according to the physical layer frame structure, performs RF front-end processing such as digital-to-analog conversion and up-conversion, and then transmits it to the wireless channel through the antenna.

[0044] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are often harsh, and the performance of traditional bit-based modulation and coding schemes degrades sharply under strong noise. This scheme directly maps semantic features to real-number symbols through a channel encoder, enabling the transmitted symbols to have stronger anti-interference capabilities at the semantic level, and providing a high-quality observation basis for subsequent semantic reconstruction based on a conditional diffusion model at the receiving end.

[0045] Specifically, the aforementioned channel coding and power normalization processes enable the transmitter to efficiently and robustly map semantic features to transmitted symbols while satisfying physical layer constraints. Compared to existing technologies, this scheme avoids the separate channel coding and modulation steps found in traditional communication, reducing processing latency and complexity. Simultaneously, it ensures a constant average transmit power for the symbol stream after power normalization, guaranteeing the system's compatibility and stability in practical deployments. Through this processing, even under low signal-to-noise ratio conditions, the receiver can still acquire a damaged symbol stream containing core semantic information, providing reliable input for subsequent semantic reconstruction based on the conditional diffusion model. This overall improves the recovery performance and reliability of the semantic communication system in complex channel environments.

[0046] S2, through the receiving end, sequentially performs channel decoding and semantic decoding on the received signal to obtain the damage depth characteristics, and obtains the current signal-to-noise ratio estimate through channel estimation.

[0047] Specifically, in the receiver processing flow, the received signal is first subjected to channel decoding and semantic decoding to recover the deep feature representation from the damaged received symbols. This process is based on the fundamental idea of ​​joint source-channel decoding, aiming to reconstruct an approximate estimate of the original semantic features from the noisy symbol stream through the inverse mapping of a deep neural network. The receiver receives the signal transmitted via a wireless channel through an antenna. This signal is superimposed with channel fading and additive white Gaussian noise during transmission. Subsequently, the received signal passes sequentially through the channel decoder and the semantic decoder. The former maps the received symbols back to the semantic feature space, while the latter performs semantic-level recovery and refinement on the mapping result, ultimately outputting the damaged deep features.

[0048] Specifically, the received signal is first input to the channel decoder, whose structure is symmetrical to the transmitting channel encoder, and is typically composed of multi-layer fully connected networks or deconvolutional networks. Let the received signal be... Its mathematical form can be expressed as: ; in, For wireless channel coefficients, which can be considered constants in quasi-static fading channels; For power is Additive white Gaussian noise; This is the real-number field symbol stream output by the transmitter. The channel decoder will receive the symbols. As input, it is mapped back to the semantic feature space through several nonlinear transformation layers, and the output is the input feature of the semantic decoder. This process can be formally represented as: ; in, The parameter set is Channel decoder, This represents the initial recovered deep features. These features are then input to a semantic decoder, whose structure is symmetrical to the sending semantic encoder. The semantic decoder typically employs a multi-head attention mechanism or a recurrent neural network to reconstruct the features at the semantic level. By capturing the internal contextual dependencies of the features, the semantic decoder refines and completes the initially recovered features, outputting the final damaged deep features. The process can be represented as follows: ; Alternatively, combining the above two formulas, it can be expressed as: ; in, This is the parameter set for the semantic decoder. That is, it serves as the conditional input in the subsequent reconstruction process of the conditional diffusion model.

[0049] Meanwhile, the receiver obtains the current signal-to-noise ratio estimate through channel estimation. In practical implementation, channel estimation and noise power measurement can be performed using pilot symbols inserted at the transmitting end. The pilot signal is inserted into the symbol stream at the transmitting end in the form of a known sequence. The receiving end estimates the channel coefficients by comparing the received pilot signal with the locally known sequence. With noise power Then, the signal-to-noise ratio estimate is calculated. This estimate reflects the current channel quality and will serve as input to subsequent channel attention coding and decoding modules, adaptively adjusting the extraction and reconstruction of semantic features.

[0050] Furthermore, the channel decoder parameter set With semantic decoder parameter set These are all learnable neural network weights, which need to be co-optimized with the corresponding module at the transmitting end during the pre-training phase or end-to-end joint training. During the pre-training phase, the loss function described in the formula can be used to independently train the semantic decoder and channel decoder, enabling them to possess basic decoding and reconstruction capabilities. In end-to-end joint training, these parameters will be further optimized to adapt to the overall system performance goals. The accuracy of channel estimation directly affects the accuracy of subsequent adaptive processing; therefore, a robust estimation algorithm must be adopted, and the pilot density must be matched with the channel change rate to obtain an accurate signal-to-noise ratio estimate.

[0051] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communications, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the received signal is often severely interfered with by noise, making it difficult for traditional decoding methods to accurately recover the original information. Through joint processing of channel decoding and semantic decoding, the receiver can initially recover deep features from the damaged symbol stream, providing an input basis for subsequent semantic reconstruction based on the conditional diffusion model. Simultaneously, the obtained SNR estimate provides crucial control information for the channel attention mechanism, enabling the receiver to adaptively adjust its reconstruction strategy according to the current channel conditions.

[0052] Specifically, through the aforementioned receiver processing, the system can achieve preliminary decoding and feature recovery of damaged signals in complex channel environments, laying the foundation for subsequent high-precision semantic reconstruction. Compared with existing technologies, this scheme, through jointly trained channel decoders and semantic decoders, enables the receiver to have stronger noise resistance, allowing it to extract damaged features containing core semantic information from received symbols even at low signal-to-noise ratios. Simultaneously, the signal-to-noise ratio estimate provided by the channel estimation module offers reliable prior information for subsequent adaptive processing, enabling the system to dynamically adjust the reconstruction process based on channel quality, thereby improving the overall recovery performance and reliability of the semantic communication system under low signal-to-noise ratio conditions.

[0053] Furthermore, S2 includes: S21, the receiver performs synchronization and channel equalization on the received symbol stream, and inputs the recovered baseband symbol sequence into the pre-trained channel decoder. The channel decoder inversely maps the baseband symbol sequence into an initial estimation vector in the semantic feature space, and inputs the initial estimation vector into the pre-trained semantic decoder. The semantic decoder adopts an attention decoding structure symmetrical to the semantic encoder of the transmitter, upsamples the initial estimation vector and reconstructs semantic relations, and outputs the damaged depth feature in the latent space that is consistent with the depth semantic feature dimension of the transmitter.

[0054] Specifically, in the receiver processing flow, synchronization and channel equalization operations are first performed on the received symbol stream. Wireless channel transmission introduces distortions such as symbol timing deviation, carrier frequency offset, and multipath fading, causing offsets in time, phase, and amplitude between the received symbol sequence and the transmitted symbol stream. The synchronization process aims to accurately estimate and correct these time-frequency parameters, enabling the receiver to correctly extract the decision point of each transmitted symbol from the continuous sample stream. Channel equalization, based on the channel state information obtained from channel estimation, compensates for inter-symbol interference caused by multipath effects, recovering a baseband symbol sequence that approximates the transmitted symbol sequence. The above preprocessing ensures that the input of the subsequent decoding module is consistent with the encoded symbols at the transmitting end in terms of statistical characteristics, providing a reliable data foundation for the inverse mapping of the deep neural network.

[0055] Furthermore, the receiver first uses the preamble or pilot symbols inserted by the transmitter for timing and carrier synchronization, for example, by employing a synchronization algorithm based on correlation operations to detect the preamble position and estimate the frequency offset. After synchronization is complete, the complete received symbol sequence is extracted. Subsequently, the channel impulse response or frequency response is estimated based on the pilot symbols, and equalizer coefficients are designed accordingly. The equalizer can be implemented using the zero-forcing criterion or the minimum mean square error criterion to perform frequency-domain or time-domain equalization on the received symbols to eliminate amplitude and phase distortions introduced by the channel. The equalized baseband symbol sequence is denoted as... Ideally, this sequence should approximate the transmitted symbol stream. The noise is superimposed, but some channel distortion and noise still remain. Then, The input is fed into the pre-trained channel decoder.

[0056] Furthermore, the channel decoder employs a deep neural network structure symmetrical to the transmitter's channel encoder, typically consisting of multiple fully connected layers or deconvolutional layers. Its function is to inversely map the equalized baseband symbol sequence from the physical layer symbol space back to the semantic feature space, outputting an initial estimation vector of the semantic features. ,Right now: ; in, This is the parameter set for the channel decoder. Although the initial estimated vector has recovered some semantic information, it is still affected by channel noise and may have lost some semantic dependencies. Therefore, it needs to be further input into the semantic decoder for refinement.

[0057] Furthermore, the semantic decoder employs an attention-based decoding structure symmetrical to the semantic encoder at the sending end, such as a Transformer-based decoder or a multi-head attention-based network. Its input is the initial estimation vector. Through upsampling operations (if dimensionality compression was performed at the sending end) and multi-layer self-attention mechanisms, the internal contextual dependencies of the features are gradually reconstructed. The attention mechanism enables the decoder to focus on the correlations between different elements in the vector, completing the semantic relationships blurred by noise, and finally outputting damaged depth features with dimensions completely consistent with the deep semantic features at the sending end. ,Right now: ; in, This is the parameter set for the semantic decoder. This serves as the conditional input in the subsequent conditional diffusion model reconstruction process. Its dimension is the same as the depth feature output by the semantic encoder at the sending end, ensuring the compatibility of subsequent processing modules.

[0058] Furthermore, the channel decoder parameter set With semantic decoder parameter set Both are learnable neural network weights. They need to be collaboratively optimized with the corresponding module at the sending end during the pre-training phase. The pre-training objective is typically to minimize the encoding / decoding reconstruction error. During the end-to-end joint training phase, these parameters will be further fine-tuned to adapt to the overall semantic reconstruction performance of the system. Output features The dimension is one of the key parameters of the system and must be strictly consistent with the dimension of the deep semantic features at the sending end to ensure dimensional matching of information before and after transmission. The dimension value can be preset as a hyperparameter according to task complexity and bandwidth constraints, such as 256 dimensions or 512 dimensions, and fixed during training.

[0059] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the received signal is often severely fading and interfered with by noise, making it difficult for traditional receivers to accurately recover bit information. Through the joint processing of channel decoding and semantic decoding, the receiver can initially recover damaged deep features in a noisy environment, providing an input basis for subsequent semantic reconstruction based on a conditional diffusion model. At the same time, the symmetrical design of the decoder structure and the transmitter ensures the consistency of the semantic feature space, enabling the receiver to fully utilize the semantic prior knowledge learned in the pre-training stage.

[0060] Specifically, the aforementioned receiver processing flow significantly improves the semantic recovery capability of the system under adverse channel conditions. Compared with existing technologies, this scheme eliminates the impact of channel distortion on decoding through synchronous and equalized preprocessing; it achieves robust inverse mapping from damaged symbols to semantic features through pre-trained channel decoders and semantic decoders; and it preserves the global dependencies of semantic features through a symmetric attention decoding structure, ensuring that the output features and the sender features maintain a high degree of semantic consistency. These characteristics provide high-quality conditional input for subsequent fine reconstruction based on the conditional diffusion model, thereby enhancing the overall noise resistance and reconstruction accuracy of the semantic communication system.

[0061] S22, the receiver uses the pilot symbols and reference signals embedded in the received signal to perform channel quality measurement, estimates the instantaneous signal-to-noise ratio of the current channel using the least squares and least mean square error criteria, generates the current signal-to-noise ratio estimate, and then uses the damage depth feature and the current signal-to-noise ratio estimate as the common input of the subsequent conditional diffusion model and channel attention decoder, so that the receiver reconstruction process depends on both the semantic content of the signal itself and the statistical characteristics of the channel state.

[0062] Specifically, in the receiver processing flow, channel quality is first measured using pilot symbols or reference signals embedded in the received signal. Pilot symbols, as prior information known to both the transmitter and receiver, carry the combined effects of channel fading and noise superposition in their received version. By comparing the received pilot symbols with the locally stored original pilot symbols, the instantaneous channel gain and noise power can be estimated, and the current signal-to-noise ratio (SNR) can be calculated. This estimation result reflects the instantaneous state of channel quality and is a key basis for subsequent adaptive processing modules to make dynamic adjustments. Using the SNR estimate and the impairment depth characteristics as inputs to subsequent modules enables the reconstruction process to simultaneously utilize the semantic content of the signal itself and the statistical characteristics of the channel state, achieving channel-aware adaptive semantic reconstruction.

[0063] Furthermore, the receiving end first extracts the sampled value at the pilot symbol location from the received signal. Let the transmitted pilot symbol be... The received pilot symbol is Under quasi-static channel conditions, satisfy ,in For channel coefficients, The noise is additive white Gaussian noise. Based on the least squares criterion, the estimated channel coefficients can be expressed as: Furthermore, if the minimum mean square error criterion is adopted, the channel estimation can be expressed as: ,in This is the cross-correlation matrix between the channel and the received pilot. This is the autocorrelation matrix of the received pilot. After obtaining the channel coefficient estimate, the noise power can be calculated by the difference between the energy of the received pilot and the energy of the estimated channel coefficient, i.e. The instantaneous signal-to-noise ratio estimate is thus calculated as follows: This estimate can be expressed in scalar form or converted to a decibel value in the logarithmic field, which facilitates subsequent network processing.

[0064] Furthermore, after obtaining the signal-to-noise ratio estimate, the receiver outputs the damage depth feature from the aforementioned steps. Compared with the current signal-to-noise ratio estimate They are processed jointly. In the specific implementation, both can be used as common inputs to the subsequent conditional diffusion model and channel attention decoder. For the conditional diffusion model, the signal-to-noise ratio estimate can be used as auxiliary conditional information, along with the damage depth feature. After concatenation, the data is input into the network, or the generation process of the diffusion model is guided by a feature modulation layer. For the channel attention decoder, the signal-to-noise ratio estimate can be used as a control factor to adjust the activation response of each layer of the decoder through scaling or biasing operations, enabling the decoding process to dynamically adjust the reconstruction strategy according to the channel quality. This fusion approach ensures that the reconstruction process at the receiver no longer depends solely on the damaged features themselves, but also incorporates statistical information about the channel state, thereby achieving adaptive adjustment when channel conditions change.

[0065] Furthermore, the accuracy of the signal-to-noise ratio (SNR) estimation directly affects the performance of subsequent adaptive modules. Excessive estimation error can introduce erroneous control information, thus reducing the reconstruction effect. Therefore, it is necessary to ensure that the density and power settings of the pilot symbols meet the estimation accuracy requirements. For example, pilots can be reasonably distributed in the time-frequency two-dimensional resources, and higher-order modulated pilots can be used to improve the estimated SNR. The SNR estimate, as the input to the subsequent neural network, usually needs to be normalized, for example, mapped to the [0,1] interval or standardized to zero mean and unit variance, to ensure the stability of network training. Damaged depth features The dimensions of the feature and the signal-to-noise ratio (SNR) estimate (usually 1-dimensional) need to be adapted during splicing. This can be achieved by using copy expansion or embedding layers to map the SNR to a higher-dimensional space to match the feature dimension.

[0066] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communications, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are time-varying and often severe, making it difficult for reconstruction models relying solely on fixed parameters to adapt to channel variations. By incorporating real-time estimated SNR into the reconstruction process, the system can dynamically adjust its semantic reconstruction strategy based on the current channel quality, preserving more details when the channel is good and focusing on restoring the semantic core when the channel is poor, thereby maintaining stable reconstruction performance across the entire SNR range.

[0067] Specifically, the above processing significantly improves the system's adaptability in time-varying channel environments. Compared with existing technologies, this scheme elevates channel state information from an auxiliary adjustment parameter to a conditional input parallel to semantic features, allowing the reconstruction process to benefit from both the semantic content of the signal and the statistical characteristics of the channel. Guided by signal-to-noise ratio (SNR) conditions, the conditional diffusion model can more accurately recover the original semantics from damaged features; the channel attention decoder dynamically adjusts the decoding weights according to the SNR, avoiding excessive noise amplification under low SNR conditions. This dual-conditional mechanism makes the receiver reconstruction process more robust and adaptable, providing key technical support for the reliable operation of semantic communication systems in complex channel environments.

[0068] S3, input the original message and the current signal-to-noise ratio estimate together into the channel attention encoder to generate a target semantic representation modulated by the channel state.

[0069] Specifically, after the receiver completes the extraction of damaged depth features and signal-to-noise ratio (SNR) estimation, the transmitter synchronously performs channel attention coding to further enhance the system's adaptability to channel conditions. The principle behind this step is that the transmission performance of a semantic communication system is directly affected by channel quality. Under low SNR conditions, if the transmitter uses a fixed coding strategy, the reconstruction difficulty at the receiver increases significantly. Therefore, this scheme introduces a channel attention mechanism, inputting the original message and the current SNR estimate into the channel attention encoder. This allows the coding process to perceive the channel state and dynamically modulate the semantic representation accordingly. This processing aims to generate a target semantic representation that adapts to channel quality, enabling the subsequent noise addition and denoising processes of the conditional diffusion model to match the channel conditions, thereby improving overall reconstruction performance.

[0070] Furthermore, the channel attention encoder is constructed using a deep neural network. Its input consists of two parts: the original message itself and the current signal-to-noise ratio (SNR) estimate obtained by the receiver through pilot estimation. The original message is first converted into a continuous vector representation through an embedding layer, while the SNR estimate is mapped to a modulation vector matching the semantic feature dimensions through a linear transformation or embedding layer. Subsequently, the two interact through a feature fusion module. This module can be designed as a multi-head attention-based structure, where the embedded representation of the original message serves as the basis for queries and key-value pairs, the SNR estimate serves as additional contextual information, and the channel state is used to calculate the weighting coefficients of each dimension in the semantic representation through an attention mechanism. The weighted semantic representation is further transformed nonlinearly through a feedforward network, ultimately outputting the target semantic representation of channel state modulation. This process can be formally represented as: ; in, This is the original message. This is the current signal-to-noise ratio estimate. The parameter set is Channel attention encoder, This is the target semantic representation for the output. In actual implementation, It can first be normalized, and then extended to the same level through a fully connected layer. The embedding dimension is consistent, and then with The embedded vectors are added or concatenated element by element, and finally relabeled by the attention layer.

[0071] Furthermore, the parameter set of the channel attention encoder Optimization is needed during end-to-end joint training. The training objective is to minimize the final reconstruction loss, enabling the network to adaptively adjust the encoding strategy of the semantic representation based on the signal-to-noise ratio estimate. Target semantic representation The embedding dimension is usually consistent with the embedding dimension of the original message, or adapted according to the input requirements of the subsequent conditional diffusion model. The accuracy of the signal-to-noise ratio (SNR) estimate directly affects the modulation effect; therefore, it is necessary to ensure the accuracy of the estimation module and simulate different SNR conditions during training so that the encoder learns modulation rules with strong generalization ability.

[0072] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communications, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are time-varying and often harsh. If the transmitter adopts a fixed coding strategy, it is difficult to balance transmission efficiency and reconstruction quality under different SNRs. By incorporating the real-time estimated SNR into the coding process, the transmitter can dynamically adjust the focus of the semantic representation according to the channel quality, retaining more detailed information when the channel is good and highlighting the semantic core when the channel is poor, thereby providing the receiver with a target representation that matches the channel conditions for conditional diffusion reconstruction.

[0073] Specifically, the aforementioned channel attention coding significantly improves the system's adaptability to channel changes. Compared to existing technologies, this scheme moves channel state information from the auxiliary adjustment parameters at the receiver to the coding stage at the transmitter, enabling the semantic representation to incorporate channel quality awareness from the very beginning of its generation. This channel-aware semantic representation serves as the initial sample for the subsequent conditional diffusion model, allowing the noise addition and denoising processes in the diffusion process to work in conjunction with channel conditions, avoiding the unrecoverable degradation characteristics that fixed coding produces under low signal-to-noise ratios. Ultimately, this mechanism effectively enhances the system's robustness in complex channel environments, laying a solid foundation for high-quality reconstruction of semantic communication.

[0074] Furthermore, S3 includes: S31, the receiver performs numerical normalization on the current signal-to-noise ratio (SNR) estimate obtained through channel estimation, maps the current SNR estimate to a feature space dimension that is the same as and can be aligned with the semantic embedding representation of the original message, and generates an SNR embedding vector. Then, the semantic embedding representation of the original message and the SNR embedding vector are concatenated along the feature dimension and fused element-wise with weights to form a joint representation containing semantic content and channel state information. The joint representation is input to the channel attention encoder, which includes a multi-head cross-attention layer. The joint representation is used as the query, and the semantic embedding representation is used as the key and value to calculate the SNR condition-guided attention weight distribution.

[0075] Specifically, in the transmitting end processing flow, to further enhance the system's ability to perceive and adapt to channel conditions, the current signal-to-noise ratio (SNR) estimate obtained by the receiving end through channel estimation is first normalized. As a continuous scalar reflecting channel quality, the original SNR estimate dynamically changes with channel conditions; for example, it may approach zero under extremely low SNR, while reaching a higher value under ideal channel conditions. Directly inputting the original value into the neural network may lead to gradient instability or model convergence difficulties during training. Normalization maps the SNR estimate to a uniform numerical range; for example, by compressing it to the [0, 1] interval through min-max normalization, or by standardizing it to make it conform to a normal distribution with zero mean and unit variance, effectively improving the stability and generalization ability of network training.

[0076] Furthermore, the normalized signal-to-noise ratio (SNR) estimate needs to be mapped to a feature space dimension that is identical and alignable to the original message semantic embedding representation. This process is achieved through an embedding layer, which can consist of a single or multiple fully connected networks. This layer transforms the normalized scalar SNR value into a dense vector of a predetermined dimension, i.e., the SNR embedding vector. Let the normalized SNR estimate be... The output of the embedding layer is Its dimension is denoted as It needs to be compared with the semantic embedding representation dimension of the original message. Maintain consistency to ensure that dimensions can be aligned in subsequent fusion operations.

[0077] Furthermore, the semantic embedding representation of the original message is generated by an independent embedding layer. The original message is first segmented into words or sub-words and converted into a token sequence. Each token is mapped to a continuous vector through an embedding lookup table, forming a token embedding sequence. If a fixed-length representation is required, average pooling or using tokens of a specific category can be used to aggregate them into a single vector, denoted as . Subsequently, semantic embedding representations are used. With signal-to-noise ratio embedding vector Fusion is performed along the feature dimension. Fusion methods may include concatenation operations, i.e. Alternatively, element-wise weighted fusion can be performed, for example, by calculating weighting coefficients through a gating mechanism. Then generate The above fusion operation forms a joint representation containing semantic content and channel state information, which serves as the input for subsequent attention coding.

[0078] Furthermore, the joint representation is input to a channel attention encoder, the core of which contains a multi-head cross-attention layer. In the attention computation mechanism, the joint representation serves as the query, and the original semantic embedding representation serves as the key and value. Specifically, let the query matrix... By joint characterization Generated by linear transformation, the key matrix Sum matrix All are represented by the original semantic embedding. Generated through different linear transformations. The multi-head attention mechanism divides the query, key, and value into multiple heads, and independently computes scaled dot product attention within each head: ; in, This is the dimension for each head. This calculation allows the signal-to-noise ratio (SNR) information in the joint representation to guide the attention mechanism, focusing on the parts of the original semantic embedding most relevant to the current channel conditions. For example, under low SNR conditions, attention weights might be more focused on core semantic components; under high SNR conditions, more weight might be allocated to detailed information. The outputs of each head are concatenated and then linearly transformed to obtain the final channel state modulation semantic representation. .

[0079] Specifically, the dimensions of the embedding layer These are important hyperparameters and need to be set according to the trade-off between model capacity and task complexity. Larger dimensions can accommodate richer modulation information but increase computational overhead. Signal-to-noise ratio normalization statistics (such as minimum, maximum, mean, and variance) need to be obtained statistically from the training set and fixed for use during inference. The number of heads in a multi-head attention layer is typically set to 8 or 16, and the dimension of each head... All the above parameters (embedding layer weights, attention layer weights, etc.) are optimized through gradient descent during end-to-end joint training.

[0080] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communications, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions change drastically and are difficult to predict. If the transmitter adopts a fixed coding strategy, it will be difficult to adapt to the optimal transmission requirements under different SNR conditions. By deeply fusing the normalized SNR embedding vector with the semantic representation and performing multi-head cross-attention modulation, the transmitter can dynamically adjust the coding focus of the semantic representation, enabling the subsequent conditional diffusion model to obtain a target representation that matches the channel conditions during reconstruction at the receiver.

[0081] Specifically, the above processing significantly enhances the system's adaptive modulation capability to channel states. Compared with existing technologies, this scheme embeds signal-to-noise ratio (SNR) information and fuses it with semantic representation in multiple dimensions, making the channel state no longer an isolated auxiliary parameter, but deeply involved in the construction process of semantic representation. The multi-head cross-attention mechanism further enables the selective enhancement or suppression of semantic features by channel conditions, making the generated modulated semantic representation more robust under low SNR and more faithful under high SNR. This provides initial samples for the subsequent noise addition and denoising process of the conditional diffusion model, co-optimizing with channel conditions, thereby improving the overall reconstruction accuracy and reliability of the semantic communication system in complex channel environments.

[0082] S32, according to the attention weight distribution, the semantic embedding representation of the original message is adaptively weighted and modulated so that the components of the semantic features that are sensitive to channel noise are suppressed and enhanced to different degrees. The target semantic representation is output after explicit calibration of the channel state. The target semantic representation is used as the initial sample in the forward diffusion process of the subsequent conditional diffusion model during the training phase, and as the ideal reconstruction target to be approximated in the reverse diffusion process.

[0083] Specifically, in the channel attention encoder's processing flow, the semantic embedding representation of the original message is adaptively weighted and modulated based on the attention weight distribution calculated by the multi-head cross-attention layer. Under different channel conditions, the sensitivity of each dimension component in the semantic features to noise varies significantly. Some feature dimensions carry core semantic information and need to be protected during transmission; while other dimensions may contain detailed or redundant information, which is easily submerged by noise under low signal-to-noise ratio conditions. Treating them equally without distinction may introduce reconstruction errors. By differentially modulating the semantic features through the attention weight distribution, the system can dynamically adjust the enhancement or suppression level of each dimension according to the current signal-to-noise ratio, achieving adaptive channel calibration of the semantic representation.

[0084] Furthermore, let the attention weight distribution of the multi-head cross-attention layer output be as follows: ,in For sequence length, The feature dimension is represented by this weight distribution, which reflects the importance of each position and dimension in the original semantic embedding representation under the current signal-to-noise ratio. Subsequently, the attention weight distribution and the original semantic embedding representation are element-wise weighted and modulated. A typical implementation is as follows: ; in, This represents element-wise multiplication. This is a representation of the original semantic embedding. This is the target semantic representation after explicit calibration based on channel conditions. This formula preserves the original semantic information through residual connections while introducing a modulation term guided by attention weights. This enhances high-importance dimensions of the features and relatively suppresses low-importance dimensions. If the attention weight distribution already incorporates awareness of channel conditions, the modulated features can adaptively highlight core semantic components and weaken details susceptible to noise interference. Modulated target semantic representation It can also be abbreviated as Its dimensions are consistent with the original semantic embedding representation, ensuring compatibility with subsequent processing modules.

[0085] Specifically, attention weight distribution The value range is typically normalized using the softmax function, ensuring that the sum of the weights at each position or dimension is 1, thereby guaranteeing the stability of the modulation process. Modulated target semantic representation. As a core input to the subsequent conditional diffusion model, it plays a dual role during the training phase. During the forward diffusion process, Using the initial sample as an example, noisy samples with different noise levels are generated by progressively adding Gaussian noise. ,Right now: ; in, These are the cumulative noise scheduling parameters. During backdiffusion, Then, as the ideal reconstruction target to be approximated, the conditional diffusion model uses the damaged depth characteristics recovered at the receiver. As a condition, from noisy samples Gradual denoising, resulting in the final reconstruction. Need to be as close as possible Its loss function is defined as: ; Specifically, through the above design, The quality of the data directly determines the target accuracy of the diffusion model training, which in turn affects the final semantic reconstruction performance of the receiver.

[0086] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communications, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions change drastically and are difficult to predict, and semantic features may be damaged to varying degrees during transmission. By dynamically modulating the semantic embedding representation through attention weight distribution, the transmitter can generate a target semantic representation that matches the current conditions based on real-time channel quality. This ensures that subsequent diffusion model training and reconstruction processes always revolve around an ideal target calibrated by the channel, thereby improving the system's adaptability to adverse channels from the source.

[0087] Specifically, the aforementioned adaptive weighted modulation mechanism significantly enhances the channel robustness of semantic representation. Compared to existing technologies, this scheme no longer treats channel state information merely as an auxiliary condition at the receiver, but rather deeply integrates it into the generation process of semantic representation at the transmitter. By explicitly calibrating semantic features through attention weight distribution, the system can achieve differentiated processing of noise-sensitive components at the feature level, enabling the transmitted semantic representation to focus more on core semantics under low signal-to-noise ratio conditions and retain richer detailed information under high signal-to-noise ratio conditions. This channel-aware semantic representation serves as the initial sample and reconstruction target for the diffusion model, ensuring that the training and inference process of the entire semantic communication system remains coordinated with channel conditions, thereby comprehensively improving the reconstruction accuracy, stability, and semantic integrity in complex channel environments.

[0088] S4. Construct a conditional diffusion model. During the reverse diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples, generating a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution.

[0089] Specifically, in the receiver processing flow, a conditional diffusion model is constructed as the core module for semantic reconstruction. The implementation principle of this model is based on the basic framework of a diffusion probability model. Its core idea is to simulate the gradual degradation process of data from its original distribution to a noisy distribution and learn an inverse denoising mapping, thereby recovering data samples conforming to the target distribution from random noise. In this invention, the conditional diffusion model is innovatively applied to semantic communication scenarios. Its forward diffusion process gradually adds noise to the target semantic representation generated by the transmitter, transforming it into pure noise; the reverse diffusion process, constrained by the damaged depth features recovered by the receiver, starts from the noisy samples and gradually denoises to generate a reconstructed semantic representation aligned with the target semantic representation in terms of semantic distribution. This mechanism enables the reconstruction process to fully utilize the residual semantic information in the damaged features, while leveraging the generation capability of the diffusion model to fill in details lost due to channel noise, thus achieving high-quality semantic recovery under low signal-to-noise ratio conditions.

[0090] Furthermore, the conditional diffusion model is constructed using a deep neural network, typically based on U-Net or Transformer architectures, and incorporates a conditional fusion module to receive impaired deep features. As a guide, during the training phase, the target semantic representation modulated by the channel attention encoder is first obtained from the transmitter. And according to the preset noise scheduling parameters Perform a forward diffusion process to generate noisy samples with different noise levels. : ; in, , This represents the total number of diffusion steps. Subsequently, the noisy samples... With current step count and conditions Common input conditional diffusion model The fusion of conditional information can be achieved in various ways, such as... After linear transformation and Concatenate along the channel dimension, or employ a cross-attention mechanism, to The representation of is a query, with The values ​​are represented as keys and values, and conditionally guided attention weights are calculated. The model output is the predicted noise component. The training loss function is defined as the mean square error between the predicted noise and the actual noise: ; Furthermore, by minimizing this loss, the model learns to perform operations under given conditions. In the case of arbitrary noise levels Accurate noise estimation enables inverse denoising capabilities. During the inference phase, from pure noise... Starting with the damaged depth characteristics recovered at the receiving end. As a condition, noise is gradually removed and iteratively generated according to the learned reverse process. Finally, the reconstructed semantic representation is obtained. The reconstruction result matches the target semantic representation of the sending end in terms of semantic distribution. The height is aligned and used as input to the subsequent channel attention decoder for final message recovery.

[0091] Specifically, diffusion steps This is a key hyperparameter affecting reconstruction quality and computational complexity, typically set to hundreds to thousands of steps. More steps result in higher generation quality but also greater inference latency. (Noise scheduling parameter) The noise addition rate is determined, and linear, cosine, or adaptive scheduling strategies can be employed. Conditional diffusion model. The network structure needs to be designed according to the dimension and characteristics of the semantic representation. If the semantic representation is in sequence form, a Transformer-based diffusion model can be used; if it is a fixed-length vector, a fully connected or convolutional U-Net can be used. The design of the conditional fusion module also affects the reconstruction performance. The cross-attention mechanism can more flexibly capture the dependency between conditional information and noisy samples, but it has a large computational cost; simple concatenation is easy to implement and suitable for resource-constrained scenarios. Model parameter set By optimizing the loss function under a large number of simulated channel conditions, the model can be generalized to different signal-to-noise ratios and noise distributions.

[0092] Specifically, this invention is particularly applicable to semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the semantic features obtained by the receiving end are often severely interfered with by noise, making it difficult to recover the original semantics through direct decoding. By constructing a diffusion model conditioned on damaged depth features, the system can utilize the prior knowledge and conditional constraints of the generative model to gradually recover a representation consistent with the original semantic distribution from the noise, effectively compensating for the performance deficiencies of traditional decoding methods under strong noise.

[0093] Specifically, the aforementioned conditional diffusion model significantly improves the accuracy and robustness of semantic reconstruction. Compared with existing generative reconstruction schemes, this scheme uses the actual recovered damage depth features at the receiver as a conditional constraint, closely linking the generation process with the transmitted content and avoiding semantic deviations caused by unconditional or weakly conditional generation. Through a multi-step denoising mechanism, the model can progressively correct distortions introduced by noise, restoring detailed information while preserving the semantic core. The final reconstructed semantic representation is highly aligned in distribution with the target semantic representation at the sender, providing high-quality input for subsequent channel attention decoding, thereby enhancing the overall reliability, integrity, and adaptability of the semantic communication system in complex channel environments.

[0094] Furthermore, S4 includes: S41, a predefined forward diffusion process is defined. Using the target semantic representation output by the channel attention encoder as the initial sample, Gaussian noise is iteratively added to the initial sample according to a predefined variance scheduling table, generating a Markov chain from clean samples to a pure noise distribution; the noisy sample at step t is represented as: ; in, For the target semantic representation, For cumulative noise dispatch coefficient, This is the sample with added noise at step t.

[0095] Specifically, in constructing the conditional diffusion model, a forward diffusion process needs to be predefined to provide a supervisory signal for subsequent backward denoising training. This process is based on the fundamental assumption of the probabilistic diffusion model: by progressively adding noise to the data, the original data distribution gradually evolves into a standard Gaussian distribution. This process is modeled as a Markov chain, where each transition depends only on the current state, and the transition kernel is defined as a Gaussian distribution with a fixed variance. Through this progressive destruction, the model can learn the inverse mapping from the noise to recover the data during the backward process, thus enabling it to generate samples that conform to the original data distribution. In this invention, the forward diffusion process uses the target semantic representation output by the channel attention encoder as the initial sample. This representation integrates semantic content and channel state information, making it an ideal target for subsequent reconstruction.

[0096] Furthermore, a variance scheduling table is first predefined, which defines the variance of the noise added at each step from step 1 to step T. Common scheduling strategies include linear scheduling, cosine scheduling, or signal-to-noise ratio-based scheduling. Linear scheduling increases the variance linearly from a small value to close to 1, while cosine scheduling makes the noise addition rate relatively gradual in the middle steps. Let the initial sample be the target semantic representation. Its dimensions are consistent with the original semantic embedding representation. For each step According to the following formula from generate : ; in, For the first The noise scheduling coefficient for each step is typically a positive number slightly less than 1. Through recursion, the noise scheduling coefficient from the initial sample to the next step can be obtained. Closed-form expression for step samples: ; in, The cumulative noise scheduling coefficient represents the distance from the initial sample to the... The proportion of signal retained during the step. When Approaching hour, Approaching 0, By approximating standard Gaussian noise, a complete Markov chain is constructed, transforming a clean sample into a purely noisy distribution. Each step in this chain involves adding noisy samples. All samples retain some of the original semantic information while being superimposed with varying degrees of noise, providing multi-scale training samples for subsequent conditional denoising networks.

[0097] Specifically, total diffusion steps It is a core hyperparameter of the forward diffusion process, typically ranging from hundreds to thousands. The more steps, the smoother the transition of the Markov chain, and the lower the learning difficulty of the inverse denoising network; however, the computational cost of training and inference also increases accordingly. Noise scheduling coefficient The rate at which noise is added is determined and needs to be optimized based on data characteristics and task requirements. Cumulative coefficient. It is a quantitative metric that measures the degree of signal preservation. During training, it is typically measured by the number of random sampling steps. This allows the network to learn denoising capabilities at different noise levels. All the scheduling parameters mentioned above are fixed before training and remain unchanged throughout the training process to ensure the determinism of the forward feed.

[0098] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the semantic features obtained at the receiving end are often severely interfered with by noise, making it difficult to directly recover the original semantics. By predefining the forward diffusion process, the system can generate a complete degradation path from clean semantic representation to pure noise at the sending end, providing the conditional diffusion model at the receiving end with a training signal that matches the real channel noise distribution, enabling the model to learn the ability to gradually recover semantics from any noise level.

[0099] Specifically, the definition and implementation of the aforementioned forward diffusion process lay a solid theoretical and data foundation for training the conditional diffusion model. Compared with existing generative reconstruction schemes, this scheme, through the progressive noise addition mechanism of a Markov chain, enables the model to learn the destructive patterns of noise on semantic features at different scales, thereby achieving accurate step-by-step repair in the reverse process. (Cumulative noise scheduling coefficient) The introduction of this feature enables the model to adaptively adjust the denoising intensity based on the noise level of the current step, improving the stability and accuracy of the reconstruction. Ultimately, the design of this forward diffusion process ensures that the conditional diffusion model can generate reconstruction results that are highly aligned in distribution with the semantic representation of the target at the sending end in complex channel environments, providing key technical support for the high-reliability transmission of semantic communication systems.

[0100] S42, construct a conditional denoising network. The conditional denoising network takes the noisy sample of the current step, the index of the current time step, and the damaged depth features as input. Through cross-layer connections and attention modulation, the damaged depth features are embedded into each decoding layer of the network, so that the denoising process is explicitly guided by the semantic content actually recovered by the receiver.

[0101] Specifically, the core of constructing the conditional diffusion model lies in designing a conditional denoising network that can effectively utilize damaged depth features as a generation guide. The implementation principle of this network is based on the fundamental idea of ​​conditional generation models, which, in the reverse denoising process, not only relies on the distribution characteristics of the noisy samples themselves but also introduces additional conditional information to constrain the generation direction, ensuring that the reconstructed result maintains semantic consistency with the specific input. In this invention, this conditional information is the damaged depth feature actually recovered at the receiving end, which carries the semantic content remaining after channel transmission. By embedding this feature into each decoding layer of the denoising network, the model always uses the actually received semantic information as a reference during the progressive denoising process, thereby generating a reconstructed result that is distributionally aligned with the target semantic representation at the sending end.

[0102] Furthermore, the conditional denoising network is constructed using a deep neural network, typically based on the U-Net architecture. This architecture achieves multi-scale feature extraction and reconstruction through an encoder-decoder structure, and directly transmits fine-grained features from the encoder to the corresponding decoder layer through cross-layer connections to preserve spatial details. The network input consists of three parts: the noisy sample from the current step... Current time step index and characteristics of damage depth The time-step index is converted into a temporal embedding vector through sinusoidal position encoding or learnable embedding, and then concatenated with the noisy samples along the channel dimension or modulated through an adaptive normalization layer. Damaged depth features. The embedding is achieved through an attention modulation mechanism. Specifically, in each layer of the decoder, the feature map of the noisy sample after processing by that layer is used as the query, and the damaged depth features are... After linear transformation, the values ​​are used as keys and values, and conditionally guided attention weights are calculated through multi-head cross-attention. This weight distribution reflects the correlation between each location in the noisy sample and each component in the damaged depth feature at the current scale. Then, conditional information is injected into the decoded feature through weighted summation. This process can be formally represented as: ; in, This represents the decoder layer index. By combining cross-layer connections with attention modulation, damaged depth features can impose constraints on the denoising process at multiple scales, ensuring that the reconstruction result conforms to both the prior distribution of the diffusion model and the semantic content actually acquired by the receiver.

[0103] Furthermore, the parameter set of the conditional denoising network is denoted as... The aforementioned diffusion loss function needs to be optimized. Network depth and the number of channels are key hyperparameters affecting model capacity and reconstruction quality, and a trade-off must be made based on the dimensionality of the semantic representation and the task complexity. The number and location of cross-layer connections usually correspond to the encoder-decoder hierarchy, ensuring that each decoder layer can obtain fine-grained information from the corresponding encoding layer. The number and dimension of attention modulation heads need to match the number of channels in the feature map to capture rich conditional dependencies. The embedding dimension of the time step index is usually set to match the number of channels in the feature map so that modulation can be performed through element-wise operations. During training, the number of random sampling steps is used. This enables the network to learn the ability to denoise using conditional information at different noise levels.

[0104] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, although the damaged depth features obtained by the receiving end contain noise, some semantic information still remains. By conditionally embedding this feature into the denoising network, the system can fully utilize this residual information during the reconstruction process, guiding the generation process towards the correct semantic direction and avoiding semantic deviation caused by relying entirely on the prior distribution of the diffusion model.

[0105] Specifically, the design of the aforementioned conditional denoising network significantly improves the consistency between the reconstructed result and the transmitted semantics. Compared with existing unconditional or weakly conditional generative reconstruction schemes, this scheme deeply integrates damaged depth features into each decoding layer of the denoising network through cross-layer connections and attention modulation, enabling conditional information to constrain the generation process at multiple scales. This multi-scale conditional guidance mechanism ensures that even in strong noise environments, the reconstruction result maintains a high degree of alignment with the original semantics, avoiding the arbitrariness and uncontrollability of the generated content. Ultimately, this design enables the conditional diffusion model to achieve high-precision semantic reconstruction under complex channel conditions, providing key technical guarantees for the reliability and integrity of semantic communication systems.

[0106] S43, during the backdiffusion process, starting from random noise and the maximum noise step sample obtained through forward diffusion, with the damaged depth feature as a fixed condition, the fixed-condition denoising network is called to iteratively predict the noise component added in the current step, gradually removing noise and restoring the semantic structure, generating an intermediate state semantic representation sequence; the semantic representation output by the backdiffusion termination step is used as the reconstructed semantic representation, and the reconstructed semantic representation meets the preset similarity threshold with the target semantic representation corresponding to the original message of the sending end in the semantic embedding space, realizing the probability alignment from damaged semantic features to complete semantic distribution.

[0107] Specifically, in the backdiffusion process, the conditional diffusion model performs a stepwise denoising reconstruction from noise to a semantic representation. This process is based on the inverse Markov chain of the diffusion probability model, with the core objective of gradually recovering the target sample that conforms to the original data distribution from a pure noise distribution through iterative denoising. In this invention, the backdiffusion process uses random noise or the sample with the maximum noise step obtained through forward diffusion as the initial state, and the damaged depth features recovered at the receiver as fixed constraints. A pre-trained conditional denoising network iteratively predicts the noise component added at each step and gradually removes it, ultimately generating a reconstructed semantic representation that is distributionally aligned with the target semantic representation at the transmitter. This mechanism combines the prior capabilities of the generative model with actual observations of channel transmission, ensuring that the reconstruction result conforms to the statistical laws of the semantic data while remaining faithful to the semantic information remaining during transmission.

[0108] Furthermore, the reverse diffusion process begins in the first... Step sample The sample is standard Gaussian noise. This can also be the sample of the largest noise step stored during the forward diffusion process. For time steps... from Decrease to 1, and perform the following iterative operation: with the current step sample Current time step index and the characteristics of damage depth under fixed conditions As input, the conditional denoising network is invoked. Predict the noise component added in the current step. Subsequently, based on the predicted noise from Remove the component and recover the sample from the previous time step. This denoising step can be implemented based on the sampling formula of the diffusion model, typically in the form of: ; in, For noise dispatch coefficient, For the cumulative scheduling coefficient, For posterior variance, An optional random term is added to increase generation diversity. By iteratively performing the above operations, a sequence of intermediate state semantic representations is generated. .when When it decreases to 0, the output is... As the final reconstructed semantic representation This represents the target semantic representation in the semantic embedding space corresponding to the original message sent by the sender. Meeting a preset similarity threshold, such as through cosine similarity or Euclidean distance, ensures that the two are highly consistent at the semantic level. This process achieves the goal of resolving semantically damaged features. Probability alignment to the complete semantic distribution, that is, mapping noisy and incomplete observations back to a clean semantic space through the conditional generation capability of the diffusion model.

[0109] Furthermore, the number of back diffusion steps Consistent with the training phase, the number of steps is typically set to several hundred to several thousand. More steps result in a more refined reconstruction process, but also increase inference latency; a trade-off must be struck based on real-time requirements. Random terms in the denoising process. The degree of introduction can be determined by parameters. Control, if the goal is to achieve deterministic reconstruction, can be set as follows: This involves employing a denoising diffusion implicit model sampling method. The similarity threshold is a key indicator for evaluating reconstruction quality and needs to be set according to the specific task. For example, in text semantic communication, the threshold can be set based on the BLEU score or semantic similarity model, while in image tasks, it can be set based on peak signal-to-noise ratio or structural similarity index. Conditional denoising network The parameters are fixed during the training phase, and only forward propagation is performed during the inference phase, without the need for gradient calculation.

[0110] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the semantic features obtained by the receiving end are often severely interfered with by noise, making it difficult for traditional decoding methods to recover the original semantics. By using a reverse diffusion process to gradually denoise based on the damaged depth features, the system can fully utilize the residual semantic information while leveraging the generative capabilities of the diffusion model to fill in the details lost due to noise, achieving high-quality reconstruction from noisy observations to clean semantics.

[0111] Specifically, the aforementioned back-diffusion process significantly improves the accuracy and reliability of semantic reconstruction. Compared with existing generative reconstruction schemes, this scheme uses the damaged depth feature as a fixed condition throughout the entire denoising process, ensuring that the generated result is always constrained by the actual transmitted content, thus avoiding the semantic drift problem caused by unconditional generation. Through an iterative denoising mechanism, the model can gradually correct the distortion introduced by noise, restoring detailed information while maintaining the semantic core, and finally generating a reconstruction result highly aligned with the target semantic representation of the sender. This probabilistic alignment mechanism ensures that the reconstructed semantics are consistent with the original message in distribution, thereby enhancing the robustness, integrity, and adaptability of the semantic communication system in complex channel environments.

[0112] S44, During the training phase, the diffusion model loss function is constructed with the objective of minimizing the mean square error between the noise predicted by the conditional denoising network and the actual Gaussian noise: ; in, For parameter set Conditional denoising network, The damaged depth characteristics recovered by the receiving end.

[0113] Specifically, during the training phase of the conditional diffusion model, the core optimization objective is formalized as minimizing the mean square error between the noise predicted by the conditional denoising network and the actual Gaussian noise. The implementation principle of this loss function is based on the training paradigm of the diffusion probability model. Its core idea is to enable the denoising network to accurately predict the added noise component from noisy samples of arbitrary noise levels through supervised learning. Since the added noise in the forward diffusion process is known to follow a standard Gaussian distribution, the task of the backward denoising network is to reverse this noisy process, gradually recovering the original sample by predicting the noise. In this invention, the loss function further introduces conditional constraints, ensuring that the denoising network's predictions not only depend on the noisy sample itself but are also guided by the damaged depth features recovered at the receiving end, thereby ensuring consistency between the generation process and the transmitted semantic content.

[0114] Furthermore, the loss function of the diffusion model is defined in the form of mathematical expectation: ; in, The target semantic representation output by the channel attention encoder at the transmitting end. For the time step index of random sampling from a uniform distribution, To and Gaussian noise of the same dimension The damaged depth characteristics recovered by the receiving end. According to the forward diffusion formula The generated noisy sample. The parameter set is A conditional denoising network whose input is noisy samples. Time step index and conditional features The output is the predicted noise component. During training, random sampling is used. , , and Calculate the L2 loss between the predicted noise and the actual noise, and then backpropagate to update the network parameters. The design of this loss function enables the model to generalize to arbitrary noise levels and learn the ability to perform accurate denoising using conditional information under different degrees of degradation.

[0115] Furthermore, the expectation in the loss function is approximated by Monte Carlo sampling, that is, a batch is randomly sampled from the training dataset in each training iteration. Correspondingly, pairs are extracted from the feature set of the receiving end. and random sampling and Time step The sampling distribution is usually uniform, that is... This ensures the model is adequately trained across all noise levels. The range of the loss function is affected by the feature dimension and noise amplitude, typically converging to a smaller value. During training, the loss change on the validation set needs to be monitored to prevent overfitting, and an early stopping mechanism can be used to determine the optimal model. The parameter set of the conditional denoising network. Optimization can be achieved using gradient descent. The optimizer can be Adam or AdamW, and the learning rate needs to be tuned according to the model size and the amount of data.

[0116] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the depth of damage features obtained by the receiving end... It carries residual semantic information after transmission through the channel, while the target semantic representation This represents the clean semantics that should ideally be recovered. Through optimization of the loss function described above, the conditional denoising network learns from noisy observations... and noisy samples The inverse mapping to clean semantics enables the model to gradually recover high-quality semantic representations based on the received damaged features in actual deployment.

[0117] Specifically, the design of the loss function in the aforementioned diffusion model provides a solid optimization foundation for achieving high-precision semantic reconstruction. Compared with existing generative reconstruction schemes, this scheme explicitly introduces the damaged depth features as a condition into the loss function, closely linking the training objective of the denoising network with the transmitted semantic content, thus avoiding semantic deviation problems that may be caused by unconditional generation. The choice of mean squared error loss ensures a precise numerical match between the predicted noise and the real noise, thereby guaranteeing the accuracy of the back-diffusion process. By randomly sampling time steps... The model has mastered denoising capabilities at each stage from slight degradation to severe degradation, enabling it to maintain stable reconstruction performance even in real-world low signal-to-noise ratio scenarios. Finally, the optimization of the loss function allows the conditional diffusion model to effectively recover the complete semantic distribution from damaged semantic features, providing crucial technical support for the robustness and reliability of semantic communication systems.

[0118] S5, the reconstructed semantic representation and the current signal-to-noise ratio estimate are input into the channel attention decoder, and a decoding operation adapted to the channel state is performed on the reconstructed semantic representation to recover the final message.

[0119] Specifically, in the final stage of the receiver processing, the reconstructed semantic representation generated by the conditional diffusion model and the current signal-to-noise ratio (SNR) estimate obtained from channel estimation are jointly input into the channel attention decoder. The principle behind this step is that although the reconstructed semantic representation has recovered a representation aligned with the target semantic distribution from the damaged depth features through the diffusion model, it is still a vector in continuous space and needs to be further mapped back to the original message space to complete the communication task. Simultaneously, the confidence level of the reconstructed semantic representation varies under different channel conditions; in low SNR environments, some semantic dimensions may still retain uncertainty. By using the SNR estimate as a conditional input to the decoder, the decoding process can dynamically adjust the mapping strategy according to channel quality, assigning greater weight to high-confidence regions and introducing prior constraints to low-confidence regions, thereby achieving channel-adaptive semantic decoding.

[0120] Furthermore, the channel attention decoder is constructed using a deep neural network, and its structure forms a symmetrical design with the channel attention encoder at the transmitting end. The decoder's input consists of two parts: first, the reconstructed semantic representation output by the conditional diffusion model. Its dimension is the same as the target semantic representation of the sending end. The two are consistent; the first is the current signal-to-noise ratio estimate obtained by the receiver through pilot estimation. The signal-to-noise ratio (SNR) estimate is first normalized and then mapped through an embedding layer to an SNR embedding vector that matches the dimension of the reconstructed semantic representation. Subsequently, the reconstructed semantic representation and the signal-to-noise ratio (SNR) embedding vector are fused, using methods such as concatenation, element-wise addition, or gated weighting. The fused features are input to the core module of the decoder, which typically includes a multi-head attention layer and a feedforward network. The attention mechanism captures the contextual dependencies within the reconstructed semantic representation, and the SNR condition guides the allocation of importance weights for each dimension. The decoder's output layer is designed according to the task type. For text transmission tasks, a linear transformation plus a softmax function can be used to output a probability distribution over the vocabulary. The final message is then generated through maximum likelihood decoding or bundle search. For image or speech tasks, deconvolutional networks can be used to reconstruct the original signal. The entire process can be formally represented as: ; in, The parameter set is Channel attention decoder, This is the news of a final recovery.

[0121] Furthermore, the parameter set of the channel attention decoder Optimization is required during end-to-end joint training. The training objective, as shown in the formula, is to minimize the final reconstructed message. Compared with the original message The cross-entropy loss between the layers. The output dimension of the decoder depends on the size of the message space. For text tasks, the output dimension equals the vocabulary size; for image tasks, the output dimension equals the pixel space dimension. The dimension of the signal-to-noise ratio (SNR) embedding layer needs to be aligned with the feature dimension of the reconstructed semantic representation, usually set to the same value, such as 256 or 512 dimensions. The number of layers, heads, and other hyperparameters of the decoder need to be weighed according to the task complexity and computational resources. In low SNR scenarios, moderately increasing the model capacity can help improve noise resistance, but overfitting should be avoided.

[0122] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are harsh and time-varying, and the quality of the reconstructed semantic representation fluctuates with the SNR. By incorporating the SNR estimate into the decoding process, the receiver can adaptively adjust its decoding strategy based on the current channel quality. When the channel is good, it fully utilizes the detailed information in the reconstructed representation; when the channel is poor, it relies on semantic priors for robust decoding, thereby maintaining stable message recovery performance across the entire SNR range.

[0123] Specifically, the aforementioned channel attention decoding process significantly improves the accuracy and robustness of the final message recovery. Compared with existing technologies, this scheme elevates channel state information from a mere auxiliary parameter to an active adjustment factor in the decoding process, enabling the decoder to dynamically optimize the mapping path from the semantic space to the message space based on the signal-to-noise ratio (SNR). Through the deep fusion of reconstructed semantic representation and SNR embedding, the decoding process retains the high-quality semantic information recovered by the diffusion model while introducing a channel quality-aware adaptive mechanism, avoiding the error propagation problem that may occur in fixed decoding under low SNR. Ultimately, this design enables the entire semantic communication system to form a complete closed loop from feature extraction and channel-aware coding at the sender end to conditional diffusion reconstruction and adaptive decoding at the receiver end, achieving high-reliability and high-fidelity semantic information transmission in complex channel environments.

[0124] Furthermore, S5 includes: S51, the receiver normalizes the current signal-to-noise ratio (SNR) estimate and performs a linear embedding transformation to generate an SNR conditional vector aligned with the dimension of the reconstructed semantic representation. The reconstructed semantic representation and the SNR conditional vector are concatenated along the feature dimension, and the SNR conditional vector is modulated to the mean and variance of the reconstructed semantic representation through a channel-wise affine transformation to form a joint semantic representation with channel state awareness. The joint semantic representation is then input into the channel attention decoder, which includes a cross-attention module. Using the joint semantic representation as the query and the pre-stored codebook and the prototype embedding of the transmitter's semantic space as the key and value, the module calculates the SNR-guided attention weights to perform feature calibration and semantic completion on the reconstructed semantic representation.

[0125] Specifically, in the final stage of receiver processing, to further improve the channel adaptability and semantic fidelity of the decoding process, the current signal-to-noise ratio (SNR) estimate is first normalized and transformed using a linear embedding layer. The principle behind this step is that the SNR estimate, as a continuous scalar reflecting channel quality, has an original numerical range that dynamically changes with channel conditions; directly using it as network input may lead to training instability. After normalization, it is mapped to a unified numerical range and then transformed into a dense vector aligned with the reconstructed semantic representation dimension through a linear embedding layer. This allows the SNR information to deeply participate in subsequent neural network calculations in the form of features, providing prior conditions for channel state awareness during the decoding process.

[0126] Furthermore, let the current signal-to-noise ratio estimate obtained by the receiver through channel estimation be... First, the scalar value is normalized, for example, by standardizing it using the mean and standard deviation of the training set, or by compressing it to the [0,1] interval using min-max normalization. The normalized scalar value is then input into a linear embedding layer, which consists of a single or multiple fully connected networks, with the output dimension set to... It needs to be combined with the reconstruction of semantic representation. Feature Dimensions Strict alignment, i.e. The output of the linear embedding layer is the signal-to-noise ratio conditional vector. .

[0127] Subsequently, the semantic representation will be reconstructed. With signal-to-noise ratio conditional vector The vectors are concatenated along the feature dimensions to form a preliminary fusion vector. To further achieve fine modulation of the reconstructed semantics by the signal-to-noise ratio, a channel-wise affine transformation mechanism is adopted, using two independent linear layers to predict the scaling factor from the concatenated vector. and offset factor Both dimensions are Then, an affine transformation is performed on the reconstructed semantic representation: ; in, This represents element-wise multiplication. This transformation enables the signal-to-noise ratio conditional vector to dynamically adjust the mean and variance of each dimension of the reconstructed semantic representation, achieving channel-state-aware feature modulation. The modulated result... This is a joint semantic representation of channel state awareness, which integrates reconstructed semantic content with current channel quality information, providing an input basis for subsequent attention decoding.

[0128] Furthermore, the aforementioned joint semantic representation is input into the channel attention decoder, the core of which contains a cross-attention module. In the attention computation mechanism, the joint semantic representation... As a query, the key and value are either a pre-stored codebook or a prototype embedding from the sender's semantic space. The codebook or prototype embedding is a set of typical semantic representations obtained through clustering or learning during the training phase; they represent benchmarks or common semantic patterns in the semantic space. Let the codebook matrix be... ,in The codebook size is used. A key matrix is ​​generated through a linear transformation. Sum matrix ,in These are learnable parameters. Query matrix. From joint semantic representation Generated through linear transformation. Multi-head cross-attention is calculated as follows: ; Furthermore, this computation enables the signal-to-noise ratio (SNR) information in the joint semantic representation to guide the attention mechanism, retrieving the baseline pattern from the codebook or prototype embedding that is most relevant to the current semantics and best suited to the channel conditions. This allows for feature calibration and semantic completion of the reconstructed semantic representation. The attention output is further processed through residual connections and a feedforward network to obtain the final calibrated semantic features, which are then mapped to the final message through the output layer. .

[0129] Furthermore, the dimension of the signal-to-noise ratio conditional vector It must be consistent with the reconstructed semantic representation to ensure dimensionality compatibility between subsequent concatenation and affine transformation. Scaling factor in affine transformation. and offset factor These are learnable parameters, typically initialized to 1 and 0 to ensure that modulation does not affect the original representation during the initial training phase. Codebook size. This is an important hyperparameter that needs to be set according to the complexity of the semantic space and the task requirements. A larger codebook can accommodate richer semantic patterns, but it increases computational overhead. The number of attention heads is usually set to 8 or 16, and the dimension of each head is... All the above parameters are optimized using gradient descent during end-to-end joint training, with the training objective being to minimize the cross-entropy loss between the final reconstructed message and the original message.

[0130] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the reconstructed semantic representation may suffer from local distortion or semantic ambiguity due to channel noise. By introducing an SNR conditional vector for affine modulation and using codebook or prototype embedding for semantic calibration through cross-attention, the receiver can dynamically compensate for uncertainties in the reconstructed semantics based on the current channel quality and retrieve missing information from the global semantic prior, thus enabling the recovery of a final result highly consistent with the original message even in noisy environments.

[0131] Specifically, the above processing significantly improves the accuracy and robustness of the final message recovery. Compared with existing technologies, this scheme uses channel-wise affine transformation to conditionally modulate the signal-to-noise ratio (SNR) into the reconstructed semantic representation, enabling the decoding process to have explicit channel awareness. By introducing pre-stored codebooks or prototype embeddings through a cross-attention mechanism, the decoder can use global semantic priors to calibrate and complete the reconstruction results, effectively compensating for potential local errors remaining in the diffusion model reconstruction process. This dual calibration mechanism ensures that even under extremely low SNR conditions, the finally recovered message maintains semantic integrity and accuracy, providing a key technical guarantee for the high-reliability transmission of semantic communication systems in complex channel environments.

[0132] S52, the calibrated reconstructed semantic representation is passed sequentially through the semantic decoding layer and the feature inverse transformation layer, and upsampling, nonlinear mapping and discretization operations symmetrical to the feature extraction network at the sending end are performed to recover the final message sequence consistent with the original message format.

[0133] Specifically, in the final stage of processing at the receiving end, the attention-calibrated reconstructed semantic representation is sequentially passed through a semantic decoding layer and a feature inverse transformation layer, performing an inverse mapping operation symmetrical to that of the sending end's feature extraction network. The principle behind this step is that the sending end maps the original message to deep semantic features in the latent space through the feature extraction network; this process typically involves dimensionality compression and nonlinear transformation. To recover the final output consistent with the original message format from the reconstructed semantic representation, a symmetrical inverse mapping network must be constructed at the receiving end. Through upsampling, nonlinear mapping, and discretization operations, the semantic vectors in the continuous space are progressively decoded into a discrete message sequence. This symmetrical design ensures the reversibility of information during the encoding-decoding process, preserving the semantic integrity of the original message to the greatest extent possible.

[0134] Furthermore, the semantic decoding layer first receives the calibrated reconstructed semantic representation, which is currently still in vector form in continuous space, with dimensions consistent with the deep semantic feature dimensions output by the sending semantic encoder. The semantic decoding layer employs a network structure symmetrical to the sending semantic encoder; for example, if the sending end uses a multi-head attention mechanism for feature compression, the receiving end uses a multi-head attention decoding module with the same number of layers for feature expansion. Through self-attention mechanisms and feedforward networks, the semantic decoding layer gradually reconstructs the internal contextual dependencies of the features and expands the feature dimensions to match the intermediate representation before the output of the sending feature extraction network. Subsequently, the inverse feature transform layer performs the inverse operation symmetrical to the sending feature extraction network. If the sending end uses a convolutional neural network for downsampling, the inverse feature transform layer uses a combination of deconvolution or upsampling plus convolution to restore the original resolution; if the sending end uses a Transformer embedding layer, the inverse feature transform layer uses a linear transformation to map the features back to the vocabulary space. Finally, a discretization operation converts the continuous output into a final message sequence consistent with the original message format. For text messages, discretization typically involves taking the word index corresponding to the highest probability in the logits vector (i.e., argmax decoding) or using beam search to generate a better sequence. The entire process can be formally represented as: ; in, For the calibrated reconstructed semantic representation, This is the final message sequence to be recovered.

[0135] Furthermore, the parameter sets of the semantic decoding layer and the inverse feature transform layer need to be collaboratively optimized with the corresponding module at the sending end during end-to-end joint training. The number of layers and attention heads in the semantic decoding layer must be symmetrical with the semantic encoder at the sending end to ensure alignment of the feature space. The upsampling factor needs to be determined based on the downsampling factor at the sending end; for example, if the sending end performs 4x downsampling, then the receiving end needs to perform 4x upsampling. The output dimension of the inverse feature transform layer needs to match the representation of the original message. For text tasks, the output dimension is the vocabulary size; for image tasks, the output dimension is the pixel value range (e.g., 0-255). The decoding strategy in the discretization operation can be selected according to the application scenario. Greedy decoding can be used in scenarios with high real-time requirements, while beam search can be used in scenarios prioritizing quality. The beam width is usually set to 3-5. All operations must ensure numerical stability to avoid gradient vanishing or exploding affecting training.

[0136] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio scenarios, such as reliable information exchange between IoT devices, semantic message delivery in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, the finally recovered message needs to be presented in its original format so that downstream tasks can use it directly. Through symmetrical semantic decoding and feature inverse transformation, the system can accurately restore the continuous semantic vectors transmitted through noisy channels and reconstructed by a diffusion model into a message format completely consistent with the sender, ensuring the semantic integrity and format compatibility of communication.

[0137] Specifically, the aforementioned symmetric decoding and inverse transform operations achieve a complete inverse mapping from the semantic space to the message space, ensuring a closed loop for end-to-end semantic communication. Compared with existing technologies, this scheme, through symmetric network design, enables the receiver to fully utilize the structural priors of the sender's encoding process, avoiding information loss caused by encoding-decoding structure mismatch. The discretization operation ensures that the final output is an interpretable and usable concrete message, rather than an intermediate representation remaining in the continuous semantic space. This design makes the entire system ready for practical deployment; whether for text, image, or speech tasks, high-fidelity semantic information recovery can be achieved through the symmetric encoding-decoding structure, providing a complete solution for the engineering application of semantic communication technology in complex channel environments.

[0138] S53, during the joint training phase, the channel attention decoder, conditional diffusion model, and channel attention encoder undergo collaborative parameter optimization using an end-to-end loss function. This end-to-end loss function employs the cross-entropy criterion to measure the semantic deviation between the final message and the original message. ; in, The original message from the sender. The final message recovered by the receiving end. The parameter set for the feature extractor For the parameter set of the semantic encoder, For the parameter set of the channel encoder, For the parameter set of the channel decoder, For the parameter set of the semantic decoder, For the parameter set of the channel attention encoder, For the parameter set of the channel attention decoder, This is the parameter set for the conditional diffusion model.

[0139] Specifically, during the joint training phase, the channel attention decoder, conditional diffusion model, and channel attention encoder undergo collaborative parameter optimization using an end-to-end loss function. The underlying principle of this step is that a semantic communication system is a complete end-to-end information transmission link. While each module can be independently pre-trained to obtain good initial parameters, independently optimized modules may exhibit suboptimal performance when operating in series. Feature extraction, semantic encoding, and channel attention modulation at the sending end; noise interference during channel transmission; and conditional diffusion reconstruction and channel attention decoding at the receiving end—any link in this complete chain can affect the final message recovery quality. Through end-to-end joint training, the parameters of all modules are collaboratively updated under the same optimization objective. This allows the modules to adaptively cooperate with each other, compensating for cross-module dependencies that were not considered during independent training, thereby achieving globally optimal system performance.

[0140] Furthermore, the end-to-end loss function uses the cross-entropy criterion to measure the semantic deviation between the finally recovered message and the original message sent. Let the original message sent be... The final message recovered by the receiving end after the complete transmission and reconstruction process is For text-based messages, the cross-entropy loss is defined as: ; in The length of the message sequence. This means that, given the generated prefixes, the model predicts the th... The cross-entropy loss function calculates the probability of the correct word at each position. For continuous data such as images or speech, cross-entropy can be replaced by mean squared error or perceptual loss, but this invention focuses on text semantic communication, so cross-entropy is used as the core metric. This loss function calculates the difference between the final output and the original message at the discrete symbol level, and its gradient can be backpropagated to all trainable modules of the entire network, including the feature extractor. Semantic encoder Channel encoder Channel attention encoder Conditional diffusion model Channel decoder Semantic decoder Channel attention decoder This enables joint optimization of global parameters.

[0141] Furthermore, end-to-end joint training typically occurs after each module has completed independent pre-training. The pre-training phase equips each module with basic semantic representation, noise-resistant transmission, and reconstruction capabilities, while joint training builds upon this foundation for fine-tuning. During training, raw messages are randomly sampled from the training dataset. Generate through complete send-transmit-receive forward propagation The cross-entropy loss is calculated, and the gradients of the parameters of each module are computed using the backpropagation algorithm. An optimizer (such as Adam) is then used to update all parameters. This process is repeated iteratively until the loss on the validation set converges or a preset early stopping condition is met. After joint training, the parameters of all modules in the entire system are collaboratively optimized, enabling the system to maximize the recovery of the original semantic information even under low signal-to-noise ratio conditions.

[0142] Furthermore, the expectation in the end-to-end loss function is approximated by mini-batch sampling. The batch size needs to be set according to the GPU memory capacity and model size, typically 32, 64, or 128. The range of the cross-entropy loss is affected by the vocabulary size and sequence length, and usually decreases to a relatively small value after convergence. The learning rate for joint training is usually lower than that in the pre-training stage, for example, set to 0.1 times the pre-training learning rate, to avoid destroying the good initialization obtained in pre-training. During optimization, the gradient norm of each module needs to be monitored to prevent gradient explosion, which can be supplemented by gradient clipping techniques. The number of joint training epochs needs to be set according to the task complexity, typically 1 / 3 to 1 / 2 of the number of pre-training epochs, to prevent overfitting.

[0143] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions are harsh and time-varying, and independent optimization of each module cannot guarantee optimal overall performance. Through end-to-end joint training, the channel attention encoder at the transmitting end can dynamically adjust the generation strategy of the target semantic representation based on the reconstruction characteristics of the conditional diffusion model at a specific SNR; the conditional diffusion model at the receiving end can optimize its denoising process to generate a more easily decoded intermediate representation based on feedback from the channel attention decoder; and the channel attention decoder can adjust its decoding weights based on the output characteristics of the diffusion model. This collaborative optimization mechanism enables the entire system to achieve adaptive performance tuning in complex channel environments.

[0144] Specifically, the end-to-end joint training significantly improves the overall reconstruction accuracy of the system under low signal-to-noise ratio conditions. Compared with existing techniques that employ staged independent training, this scheme, through a unified cross-entropy loss function, enables all modules to co-evolve under the same objective, avoiding local optima and inter-module mismatch problems that may result from independent optimization. Joint training allows the channel attention encoder to generate a more favorable target representation for subsequent diffusion reconstruction, enables the conditional diffusion model to produce reconstruction results that are easier to decode, and enables the channel attention decoder to more accurately recover the original message. Ultimately, this co-optimization mechanism enables the entire semantic communication system to achieve global optimum in robustness, reconstruction accuracy, and semantic integrity under complex channel environments, providing a complete training solution for the engineering application of semantic communication technology.

[0145] S54, through the dynamic adjustment mechanism of the attention weights inside the channel attention decoder, enables the decoding process to adaptively switch the degree of dependence on the reconstructed semantic representation under different signal-to-noise ratio conditions. This includes that when the signal-to-noise ratio estimate is higher than the threshold, the decoder focuses on directly utilizing the detailed features in the reconstructed semantic representation; when the signal-to-noise ratio estimate is lower than the threshold, the decoder enhances its dependence on the semantic prototype embedding, suppresses the feature distortion introduced by noise, and achieves robust semantic recovery with channel state awareness.

[0146] Specifically, in the design of the channel attention decoder, a dynamic adjustment mechanism for internal attention weights is introduced, enabling the decoding process to adaptively switch the degree of dependence on the reconstructed semantic representation based on the current signal-to-noise ratio (SNR) conditions. This step is based on the direct impact of SNR conditions on the reliability of semantic features. Under low SNR conditions, residual noise may introduce feature distortion into the reconstructed semantic representation. If the decoder still over-relies on these distorted features, it will lead to errors in the final recovered message. Conversely, under high SNR conditions, the detailed information in the reconstructed semantic representation is more reliable, and directly utilizing these details helps improve recovery accuracy. By introducing an SNR-aware dynamic adjustment mechanism, the decoder can adaptively balance the dependence on the reconstructed semantic representation and the semantic prototype embedding under different channel conditions, achieving channel-state-aware robust semantic recovery.

[0147] Furthermore, the internal attention weight dynamic adjustment mechanism of the channel attention decoder is implemented through a signal-to-noise ratio (SNR) controlled gating unit. This gating unit uses the current SNR estimate... As input, after normalization and linear transformation, a scalar weight between 0 and 1 is generated. ,Right now: ; in, It is the sigmoid activation function. This is the normalized signal-to-noise ratio estimate. The weight... Used to adjust the fusion ratio of the two information paths in the decoder: one path directly utilizes the reconstructed semantic representation. Another path relies on pre-stored semantic prototype embeddings or codebooks to capture detailed features. In cross-attention calculation, the final attention output can be expressed as: ; The first term uses the reconstructed semantic representation itself or its transformation as the key and value, focusing on utilizing detailed information in the input; the second term uses the semantic prototype embedding as the key and value, focusing on utilizing global semantic priors for feature calibration. When the signal-to-noise ratio estimate is higher than a preset threshold, As the signal-to-noise ratio (SNR) approaches 1, the decoder focuses on directly utilizing detailed features from the reconstructed semantic representation to retain as much original information as possible. When the SNR estimate is below a threshold, Approaching zero, the decoder enhances its reliance on semantic prototype embeddings, retrieving the most relevant semantic patterns from the prototype library through an attention mechanism, suppressing feature distortion introduced by noise, and thus achieving robust decoding. The threshold itself can be a learnable parameter or set empirically, and a soft switch is achieved through a smooth transition using the sigmoid function, avoiding the discontinuities caused by abrupt binarization switching.

[0148] Furthermore, the linear layer weights in the gating unit are learnable parameters that are optimized together with other parameters of the decoder during end-to-end joint training. The value range is always between 0 and 1, ensuring the smoothness and stability of the fusion ratio. Semantic prototype embedding. The dimension of the prototype needs to be consistent with the feature dimension of the reconstructed semantic representation. The signal-to-noise ratio (SNR) threshold needs to be set based on the complexity of the semantic space, typically ranging from several hundred to several thousand. The SNR threshold can be initialized to an empirical value, such as the minimum acceptable operating SNR for the system, and fine-tuned during training using gradient descent. Hyperparameters such as the number of heads and dimensionality in the attention computation should be kept consistent with other parts of the decoder to ensure overall computational efficiency.

[0149] Specifically, this technical solution is particularly suitable for semantic communication systems in low signal-to-noise ratio (SNR) scenarios, such as reliable information exchange between IoT devices, semantic message transmission in emergency communication, data acquisition and transmission in remote monitoring, and task collaboration in mobile edge computing environments. In these applications, channel conditions vary drastically and are difficult to predict, causing significant fluctuations in the quality of reconstructed semantic representations with increasing SNR. By dynamically adjusting the dependence on reconstructed semantic representations and semantic prototype embedding, the decoder can fully utilize transmission details to improve recovery quality when channel conditions are favorable, and maintain basic semantic integrity by leveraging semantic prior knowledge when channel conditions are poor, thereby maintaining stable decoding performance across the entire SNR range.

[0150] Specifically, the aforementioned dynamic adjustment mechanism significantly improves the system's robustness and adaptability in time-varying channel environments. Compared with existing technologies, this scheme, through a signal-to-noise ratio (SNR)-guided soft handover mechanism, enables the decoder to dynamically balance the two information sources based on channel quality, avoiding the suboptimal performance issues that may occur with fixed-weight fusion when the SNR changes. Under high SNR conditions, the decoder retains more transmission details, improving recovery accuracy; under low SNR conditions, the decoder effectively suppresses noise-induced feature distortion by enhancing its dependence on prototype embedding, preventing the spread of errors caused by noise contamination of detailed information. This mechanism enables the entire semantic communication system to achieve channel-state-aware robust semantic recovery in complex channel environments, providing key technical support for high-reliability communication applications.

[0151] This invention provides an end-to-end semantic reconstruction method for low signal-to-noise ratio (SNR) scenarios. The method involves: At the transmitting end, deep semantic feature extraction and joint encoding of the original message generate a symbol stream suitable for channel transmission; at the receiving end, channel decoding and semantic decoding are performed on the received signal to obtain damaged depth features and channel state estimates; based on this, the original message and the SNR estimate are input into a channel attention encoder to generate a target semantic representation modulated by the channel state; then, a conditional diffusion model constrained by the damaged depth features is constructed, generating a reconstructed semantic representation aligned with the distribution of the target semantic representation through forward noise addition and reverse denoising processes; finally, the reconstructed semantic representation and the SNR estimate are input together into a channel attention decoder for adaptive decoding to recover the final message. This invention, by integrating a channel-aware attention mechanism and a conditional diffusion model, achieves integrated modeling from multi-level feature extraction and cross-layer information fusion to semantic reconstruction, significantly improving the accuracy of semantic information recovery and system robustness under low SNR conditions, and enhancing the model's adaptability and engineering applicability in complex channel environments.

[0152] Example 2 To achieve the above invention, embodiments of the present invention also provide another end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios, including: like Figure 2 As shown, Alice and Bob are the two parties in a text message transmission task. The original message is processed by the feature extractor. Obtain deep features , After semantic encoder Obtain semantic features , After encoder Obtain the symbol stream transmitted over the wireless channel. ,Right now

[0153] in, These are the parameter sets for the feature extractor, semantic encoder, and channel encoder, respectively.

[0154] Symmetrically, as the receiver, Bob receives the signal as follows: , means as follows:

[0155] in, The wireless channel coefficients between the transmitter and receiver (assumed to be constant in a quasi-static channel). For power is Additive white Gaussian noise (AWGN) can be used, and Bob can estimate the signal-to-noise ratio using pilot signals. Received Passing through the channel decoder in sequence and semantic decoder Subsequently, the restored depth features were obtained. , represented as:

[0156] in, These are the parameter sets for the semantic decoder and the channel decoder, respectively.

[0157] During the training phase, and As a channel attention encoder The input is used to obtain the coding result after considering the channel features. , represented as:

[0158] in This is the parameter set for the channel attention encoder. Subsequently, Input conditional diffusion model During training, a progressive noise-adding operation is performed to generate noisy samples at different noise levels. In the backdiffusion phase, the depth features recovered at the receiver are used. As a conditional input, the conditional diffusion model applies noisy samples. Perform progressive denoising to obtain the reconstructed semantic representation. Its relationship with signal-to-noise ratio estimation Input together into the channel attention decoder Adaptive decoding of semantic information yields the final recovered message. , represented as:

[0159] in , This is the parameter set for the channel attention decoder and the conditional diffusion model.

[0160] In the above scenario, the proposed semantic information recovery scheme aims to maximize Bob's reconstruction accuracy in order to accurately recover the required text information.

[0161] Furthermore, the basic steps of the training process include: Step 1: To support the independent use of the semantic encoder and decoder, the semantic encoder and decoder are first trained. The semantic loss function is designed as follows:

[0162] in, This refers to batch data used during the training process. This is the cross-entropy function.

[0163] Step Two: Similar to the training process described above, to support the independent use of the channel encoder and decoder modules, the channel encoder and decoder are trained, and their loss functions are designed as follows:

[0164] Through the above training, the channel coding and decoding module acquires basic noise-resistant transmission capabilities under different channel conditions.

[0165] Step 3: To improve the system's semantic reconstruction capability under low signal-to-noise ratio conditions, the conditional diffusion model is trained. During training, the semantic representation output by the channel attention encoder is used. Using the initial sample, a progressive noise-adding operation is performed to obtain the noisy sample. , represented as:

[0166] In the back-diffusion phase, the semantic features recovered at the receiver are... As a conditional input, the conditional diffusion model applies the noisy samples. The loss function for progressive denoising is defined as follows:

[0167] Step 4: To further reduce the redundancy caused by the split coding scheme, based on the pre-trained network modules, a joint loss function is designed as follows:

[0168] in These are the original message sent from the sender and the reconstructed message received from the receiver, as defined in formulas (1) and (3), respectively.

[0169] In summary, the training algorithm process is as follows: 1. Initialize the parameters of each network module; 2. Repeat training until all epochs are completed or the early stop condition is met: 2.1. Train the semantic codecs of both communicating parties independently according to formula (6); 2.2. Train the channel codecs of both communicating parties independently according to formula (7); 2.3. Train the receiver conditional diffusion model independently according to formula (9); 2.4. Jointly train the entire network module of both communicating parties according to formula (10); 3. Return the system model parameters of the legitimate communicator.

[0170] Another end-to-end semantic reconstruction method for low signal-to-noise ratio (SNR) scenarios, as described in this invention, involves the sending end extracting semantic features and jointly encoding the original message to generate a symbol stream, while the receiving end performs channel decoding and semantic decoding on the received signal to obtain the damage depth features and SNR estimate. The original message and SNR estimate are then input into a channel attention encoder to generate a target semantic representation. A diffusion model conditioned on the damage depth features is then constructed, and a reconstructed semantic representation aligned with the target semantic representation is generated through progressive denoising. Finally, the final message is recovered via adaptive decoding by the channel attention decoder. This method effectively solves the problem of reliable semantic information recovery under low SNR conditions in existing technologies. It achieves integrated modeling from multi-level feature extraction and cross-layer information fusion to semantic reconstruction, significantly improving the accuracy and robustness of semantic recovery, and enhancing the model's adaptability and engineering applicability in complex channel environments.

[0171] Example 3 To achieve the above invention, such as Figure 3 As shown, this embodiment also provides an end-to-end semantic reconstruction device 10 for low signal-to-noise ratio scenarios. The device 10 includes: The transmitting end semantic coding and transmission module 100 is used to extract semantic features from the original message through the transmitting end, generate deep semantic features to be transmitted, and sequentially perform semantic coding and channel coding on the deep semantic features to generate a symbol stream for transmission through a wireless channel. The receiver initial decoding and channel estimation module 200 is used to perform channel decoding and semantic decoding on the received signal sequentially through the receiver to obtain the damage depth features, and to obtain the current signal-to-noise ratio estimate through channel estimation. The channel state modulation module 300 is used to input the original message and the current signal-to-noise ratio estimate into the channel attention encoder to generate a target semantic representation modulated by the channel state. The conditional diffusion reconstruction module 400 is used to construct a conditional diffusion model. During the back diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples and generate a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution. The channel attention decoding and message recovery module 500 is used to input the reconstructed semantic representation and the current signal-to-noise ratio estimate into the channel attention decoder, perform a decoding operation adapted to the channel state on the reconstructed semantic representation, and recover the final message.

[0172] This invention discloses an end-to-end semantic reconstruction device for low signal-to-noise ratio (SNR) scenarios. The device uses a semantic encoding and transmission module at the transmitting end to extract features from the original message and perform joint encoding to generate a symbol stream. At the receiving end, an initial decoding and channel estimation module obtains the damage depth features and SNR estimate. A channel state modulation module inputs the original message and SNR estimate into a channel attention encoder to generate a target semantic representation. A conditional diffusion reconstruction module then generates a reconstructed semantic representation aligned with the target semantic distribution, using the damage depth features as a conditional constraint. Finally, a channel attention decoding and message recovery module performs adaptive decoding and outputs the final message. This effectively solves the problem of reliable semantic information recovery under low SNR conditions in existing technologies. It achieves integrated modeling from multi-level feature extraction and cross-layer information fusion to semantic reconstruction, significantly improving the accuracy and robustness of semantic recovery and enhancing the model's adaptability and engineering applicability in complex channel environments.

[0173] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios described above.

[0174] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios as described in the foregoing embodiments.

[0175] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0176] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. An end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios, characterized in that, include: The sending end extracts semantic features from the original message to generate deep semantic features to be transmitted, and performs semantic coding and channel coding on the deep semantic features in sequence to generate a symbol stream for transmission through the wireless channel. By performing channel decoding and semantic decoding sequentially on the received signal at the receiving end, the damage depth characteristics are obtained, and the current signal-to-noise ratio estimate is obtained through channel estimation. The original message and the current signal-to-noise ratio estimate are input together into the channel attention encoder to generate a target semantic representation modulated by the channel state. A conditional diffusion model is constructed. During the back diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples and generate a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution. The reconstructed semantic representation and the current signal-to-noise ratio estimate are input into the channel attention decoder. The reconstructed semantic representation is then subjected to a decoding operation adapted to the channel state to recover the final message.

2. The method as described in claim 1, characterized in that, The step of extracting semantic features from the original message at the sending end to generate deep semantic features to be transmitted, and then sequentially performing semantic coding and channel coding on the deep semantic features to generate a symbol stream for transmission through a wireless channel includes: The original message is input into a pre-trained feature extraction network at the sending end. The deep semantic features of the original message in the latent space are extracted through multi-layer nonlinear transformation. The deep semantic features are then input into a pre-trained semantic encoder. The semantic encoder uses a multi-head attention mechanism to model the global dependency relationship and compress the dimension of the deep semantic features, and outputs a fixed-length semantic feature vector. The semantic feature vector is input to a pre-trained channel encoder, which maps the semantic feature vector into a real-domain symbol stream suitable for physical channel transmission, applies power normalization constraints to the real-domain symbol stream, and then transmits the real-domain symbol stream to the receiver via a wireless channel.

3. The method as described in claim 1, characterized in that, The step of sequentially performing channel decoding and semantic decoding on the received signal at the receiving end to obtain the impairment depth features, and obtaining the current signal-to-noise ratio estimate through channel estimation, includes: The receiver performs synchronization and channel equalization on the received symbol stream, and inputs the recovered baseband symbol sequence into a pre-trained channel decoder. The channel decoder inversely maps the baseband symbol sequence into an initial estimation vector in the semantic feature space, and inputs the initial estimation vector into a pre-trained semantic decoder. The semantic decoder adopts an attention decoding structure that is symmetrical to the semantic encoder at the transmitter, upsamples the initial estimation vector and reconstructs semantic relations, and outputs a damaged depth feature in the latent space that is consistent with the dimension of the deep semantic feature at the transmitter. By using the pilot symbols and reference signals embedded in the received signal at the receiving end, the channel quality is measured. The instantaneous signal-to-noise ratio of the current channel is estimated by the least squares and least mean square error criteria, and the current signal-to-noise ratio estimate is generated. Then, the damage depth feature and the current signal-to-noise ratio estimate are used as the common input of the subsequent conditional diffusion model and channel attention decoder, so that the receiving end reconstruction process depends on both the semantic content of the signal itself and the statistical characteristics of the channel state.

4. The method as described in claim 1, characterized in that, The step of inputting the original message and the current signal-to-noise ratio estimate into the channel attention encoder to generate a channel state modulated target semantic representation includes: The receiver performs numerical normalization on the current signal-to-noise ratio (SNR) estimate obtained through channel estimation, mapping the current SNR estimate to a feature space dimension that is identical and alignable to the semantic embedding representation of the original message, thus generating an SNR embedding vector. Then, the semantic embedding representation of the original message and the SNR embedding vector are concatenated along the feature dimension and fused element-wise with weights to form a joint representation containing semantic content and channel state information. This joint representation is input to the channel attention encoder, which includes a multi-head cross-attention layer. Using the joint representation as the query and the semantic embedding representation as the key and value, the encoder calculates the SNR-conditionally guided attention weight distribution. Adaptive weighted modulation is applied to the semantic embedding representation of the original message according to the attention weight distribution, so that the components of the semantic features that are sensitive to channel noise are suppressed and enhanced to different degrees. The target semantic representation is output after explicit calibration of the channel state. The target semantic representation is used as the initial sample in the forward diffusion process of the subsequent conditional diffusion model during the training phase, and as the ideal reconstruction target to be approximated in the backward diffusion process.

5. The method as described in claim 1, characterized in that, The construction of the conditional diffusion model, which uses the damaged depth feature as a conditional constraint during the back-diffusion process, performs progressive denoising on the noisy samples to generate a reconstructed semantic representation that is semantically aligned with the target semantic representation, includes: A predefined forward diffusion process is used, with the target semantic representation output by the channel attention encoder as the initial sample. Gaussian noise is iteratively added to the initial sample according to a predefined variance scheduling table, generating a Markov chain from clean samples to a pure noise distribution; the noisy sample at step t is represented as: ； in, For the target semantic representation, For cumulative noise dispatch coefficient, For the noisy sample added at step t; A conditional denoising network is constructed. The conditional denoising network takes the noisy sample of the current step, the index of the current time step, and the damaged depth features as input. Through cross-layer connections and attention modulation, the damaged depth features are embedded into each decoding layer of the network, so that the denoising process is explicitly guided by the semantic content actually recovered by the receiver. During the backdiffusion process, starting from random noise and the maximum noise step sample obtained through forward diffusion, the fixed-condition denoising network is called to iteratively predict the noise component added in the current step, using the damaged depth feature as a fixed condition. The noise is gradually removed and the semantic structure is restored, generating an intermediate state semantic representation sequence. The semantic representation output at the backdiffusion termination step is used as the reconstructed semantic representation. The reconstructed semantic representation satisfies the preset similarity threshold with the target semantic representation corresponding to the original message at the sending end in the semantic embedding space, thus achieving probability alignment from damaged semantic features to complete semantic distribution. During the training phase, the diffusion model loss function is constructed with the objective of minimizing the mean square error between the noise predicted by the conditional denoising network and the actual Gaussian noise: ； in, For parameter set Conditional denoising network, The damaged depth characteristics recovered by the receiving end.

6. The method as described in claim 1, characterized in that, The step of inputting the reconstructed semantic representation and the current signal-to-noise ratio estimate into the channel attention decoder, performing a decoding operation adapted to the channel state on the reconstructed semantic representation, and recovering the final message includes: The receiver normalizes the current signal-to-noise ratio (SNR) estimate and performs a linear embedding transformation to generate an SNR conditional vector aligned with the dimensions of the reconstructed semantic representation. The reconstructed semantic representation and the SNR conditional vector are concatenated along the feature dimension, and a channel-wise affine transformation is used to modulate the SNR conditional vector to the mean and variance of the reconstructed semantic representation, forming a channel-state-aware joint semantic representation. This joint semantic representation is then input to a channel attention decoder, which includes a cross-attention module. Using the joint semantic representation as the query and the pre-stored codebook and the prototype embedding of the transmitter's semantic space as the key and value, the module calculates the SNR-guided attention weights to perform feature calibration and semantic completion on the reconstructed semantic representation. The calibrated reconstructed semantic representation is passed sequentially through the semantic decoding layer and the feature inverse transformation layer, and upsampling, nonlinear mapping and discretization operations symmetrical to the feature extraction network at the sending end are performed to recover the final message sequence consistent with the original message format. During the joint training phase, the channel attention decoder, conditional diffusion model, and channel attention encoder undergo collaborative parameter optimization using an end-to-end loss function. This end-to-end loss function employs the cross-entropy criterion to measure the semantic deviation between the final message and the original message. ； in, The original message from the sender. The final message recovered by the receiving end. The parameter set for the feature extractor For the parameter set of the semantic encoder, For the parameter set of the channel encoder, For the parameter set of the channel decoder, For the parameter set of the semantic decoder, For the parameter set of the channel attention encoder, For the parameter set of the channel attention decoder, This is the parameter set for the conditional diffusion model; Through the dynamic adjustment mechanism of the attention weights inside the channel attention decoder, the decoding process adaptively switches the degree of dependence on the reconstructed semantic representation under different signal-to-noise ratio (SNR) conditions. When the SNR estimate is higher than the threshold, the decoder focuses on directly utilizing the detailed features in the reconstructed semantic representation; when the SNR estimate is lower than the threshold, the decoder enhances its dependence on the semantic prototype embedding, suppresses feature distortion introduced by noise, and achieves robust semantic recovery with channel state awareness.

7. An end-to-end semantic reconstruction device for low signal-to-noise ratio scenarios, characterized in that, include: The sending end semantic encoding and transmission module is used to extract semantic features from the original message through the sending end, generate deep semantic features to be transmitted, and sequentially perform semantic encoding and channel encoding on the deep semantic features to generate a symbol stream for transmission through the wireless channel. The receiver initial decoding and channel estimation module is used to perform channel decoding and semantic decoding on the received signal sequentially through the receiver to obtain the damage depth features, and to obtain the current signal-to-noise ratio estimate through channel estimation. The channel state modulation module is used to input the original message and the current signal-to-noise ratio estimate into the channel attention encoder to generate a target semantic representation modulated by the channel state. The conditional diffusion reconstruction module is used to construct a conditional diffusion model. During the back diffusion process, the conditional diffusion model uses the damaged depth feature as a conditional constraint to perform stepwise denoising on the noisy samples and generate a reconstructed semantic representation that is aligned with the target semantic representation in terms of semantic distribution. The channel attention decoding and message recovery module is used to input the reconstructed semantic representation and the current signal-to-noise ratio estimate into the channel attention decoder, perform a decoding operation adapted to the channel state on the reconstructed semantic representation, and recover the final message.

8. An electronic device, comprising: processor; The memory stores executable instructions; when the processor executes the instructions, it implements an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements an end-to-end semantic reconstruction method for low signal-to-noise ratio scenarios as claimed in any one of claims 1-6.

Images