A receiving device for joint decoding and implicit equalization in a high-speed mobile OFDM system.

Abstract

This invention provides a receiving device for joint decoding and implicit equalization in a high-speed mobile OFDM system. This receiving device acquires a time-domain received signal carrying image information and reconstructs the original image based on the time-domain received signal. The receiving device includes: an OFDM demodulator for performing OFDM demodulation on the received time-domain received signal to obtain a frequency-domain received signal; a hybrid channel estimator that uses a neural network to extract the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the frequency-domain received signal, and estimates channel state information based on the dynamic fusion characteristics; and a decoder that uses a neural network to perform semantic decoding and reconstruction based on the channel state information and the frequency-domain received signal to obtain the reconstructed original image. This invention effectively resists channel estimation errors and improves semantic image reconstruction performance by jointly decoding and reconstructing using channel state information and the frequency-domain received signal.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to channel estimation and equalization in semantic communication, and more specifically to a receiving device for joint decoding and implicit equalization in a high-speed mobile OFDM system. Background Technology

[0002] In wireless communication systems, the core task of the receiver is to accurately recover the original information transmitted by the transmitter from the received signal, which is affected by noise and channel distortion (such as fading and interference). In modern communication systems (such as OFDM), this typically requires two key steps: channel estimation and channel equalization.

[0003] 1) Channel Estimation: The receiver first uses the pilot information embedded in the signal to estimate the channel response (CSI) on the time-frequency grid. Existing techniques include traditional pilot-based estimation algorithms, such as least squares (LS) or minimum mean square error (MMSE) estimation. In recent years, with the development of deep learning, some channel estimators based on neural networks (such as CNN, RNN, LSTM) have emerged, which attempt to learn the time-frequency correlation of the channel to obtain more accurate estimation results.

[0004] However, existing deep learning methods have significant limitations, including: CNN-based methods are limited by fixed convolution kernel sizes and have difficulty capturing long-range dependencies in the frequency domain. This is because multipath delay spread makes it difficult to obtain physical fading correlations between subcarriers that are far apart or even span the entire frequency band; RNN / LSTM-based methods can model temporal dynamics, but they are prone to gradient vanishing or gradient exploding when processing long sequences, and are difficult to parallelize, resulting in unsatisfactory estimation accuracy and computational efficiency in high-speed moving scenarios.

[0005] 2) Channel equalization: After obtaining channel estimation Afterward, the receiver needs to perform channel equalization to compensate for the distortion caused by the channel to the signal. Traditional methods involve performing explicit equalization in the frequency domain, such as zero-forcing (ZF) equalization or MMSE equalization. The principle behind this is to calculate the inverse of the channel matrix. To restore the launch symbol However, in high-speed mobile scenarios, the channel can change significantly within one OFDM symbol period, resulting in time-varying channel matrix characteristics. Traditional single-tap equalizers (assuming that the subcarriers are orthogonal and independent) cannot effectively suppress inter-carrier interference (ICI) introduced by time-varying channels. Furthermore, at deep fading locations (where the channel gain is close to zero), explicit equalization will significantly amplify noise, leading to severe degradation of equalization performance.

[0006] In addition to channel equalization, source decoding is also included. Source decoding: In a separated communication system, this refers to the decoding of equalized symbols... The data is then fed into a channel decoder (such as LDPC decoder or Turbo decoder), followed by a source decoder (such as JPEG decoder or HEVC decoder) to reconstruct the original data (such as an image). In an end-to-end (E2E) semantic communication (JSCC) system, the equalized symbols... The data will be directly fed into a Joint Source-Channel Decoder (JSCC Decoder) to reconstruct the data. It is worth noting that because the JSCC system omits the explicit channel coding / decoding stage, its decoder is more sensitive to the quality of the input signal (i.e., the accuracy of channel equalization and channel estimation). Any estimation or equalization error will directly affect the final semantic reconstruction quality.

[0007] The aforementioned existing technologies have serious problems and drawbacks when applied to high-speed mobile scenarios:

[0008] 1) Channel estimation mismatch under high dynamics: High-speed movement (high Doppler) leads to drastic time-varying characteristics of the channel, exhibiting complex dynamics in both the frequency domain (across subcarriers) and the time domain (across OFDM symbols). Traditional LS / MMSE estimators, based on linear assumptions, cannot effectively capture this complex nonlinear time-frequency correlation. Existing deep learning estimators (such as simple CNNs or RNNs) often focus only on modeling one dimension, either the time or frequency domain, making it difficult to simultaneously and accurately capture both global correlations in the frequency domain and local evolution trends in the time domain. Furthermore, in practical systems, pilot density is typically low (e.g., 5%-10%) to reduce pilot overhead, further exacerbating the estimation difficulty. Taking a 3.5GHz carrier and a movement speed of 50m / s as an example, the channel coherence bandwidth is approximately 200kHz, and the coherence time is approximately 0.5ms. If the pilot spacing exceeds the coherence bandwidth or coherence time, traditional interpolation methods will fail, resulting in a severe decrease in channel estimation accuracy under high dynamics, with the normalized mean square error (NMSE) potentially deteriorating by 3-5dB.

[0009] 2) The "error propagation" and "error amplification" effects caused by explicit balancing: This is a fundamental flaw of traditional cascaded architectures. In the above: first estimate... Rebalancing In the post-decoding pipeline, channel estimation There is always an error ( (), especially at high speeds and low SNR. When this inaccuracy Used for explicit equilibration (such as calculation) When this happens, the estimation error will be amplified and propagated to the recovered sign. In particular, in zero-forcing (ZF) equilibrium, if If the fading is close to zero on some subcarriers (deep fading), then This will tend towards infinity, leading to extreme amplification of noise (noise enhancement factor can reach 10-20 dB); in MMSE equalization, although the noise amplification is mitigated by introducing a regularization term, the estimation error... The error will still be nonlinearly amplified through the inverse operation of the equalization matrix, and this amplification effect is positively correlated with the condition number of the channel matrix. This amplified error will severely contaminate the decoder input, causing a sharp deterioration in the performance of subsequent source decoding or semantic reconstruction (such as image reconstruction), which cannot be remedied even if the decoder itself has excellent performance.

[0010] Therefore, when existing technologies are applied to high-speed mobile scenarios, the channel estimation accuracy is severely reduced under high dynamic conditions, which further leads to the subsequent explicit equalization amplification of errors and propagation into the recovered symbols, thereby causing a sharp deterioration in the final source decoding or semantic reconstruction performance.

[0011] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention

[0012] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a receiving device for joint decoding and implicit equalization of a high-speed mobile OFDM system.

[0013] The objective of this invention is achieved through the following technical solution:

[0014] According to a first aspect of the present invention, a receiving apparatus is provided, which, during the transmission of an original image, obtains a time-domain received signal carrying image information and reconstructs the original image based on the time-domain received signal. The receiving apparatus includes: an OFDM demodulator for performing OFDM demodulation on the received time-domain received signal to obtain a frequency-domain received signal; a hybrid channel estimator that uses a neural network to extract the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the frequency-domain received signal, and estimates channel state information based on the dynamic fusion characteristics; and a decoder that uses a neural network to perform semantic decoding and reconstruction based on the channel state information and the frequency-domain received signal to obtain the reconstructed original image.

[0015] In some embodiments of the present invention, the decoder includes: a preprocessing module for preprocessing and splicing the estimated channel state information and frequency domain received signal to obtain joint input data, including channel state data and OFDM signal data corresponding to each subcarrier in the frequency domain received signal; and a feature extraction module that uses a deep neural network to extract the channel state features and OFDM signal features of each subcarrier from the joint input data, assigns attention weights to each subcarrier according to the channel state features and OFDM signal features of each subcarrier, and performs semantic decoding and reconstruction of the OFDM signal features of each subcarrier based on the attention weights of each subcarrier to obtain the reconstructed original image.

[0016] In some embodiments of the present invention, the method of assigning attention weights to each subcarrier in the feature extraction module includes: using a cross-attention mechanism, using the OFDM signal features of each subcarrier as the query, and the channel state features of each subcarrier as the key and value, calculating the attention weights of each subcarrier and assigning them to each subcarrier; wherein, the inner product between the OFDM signal features and the channel state features corresponding to the subcarrier is positively correlated with the attention weights.

[0017] In some embodiments of the present invention, subcarriers with attention weights less than a preset threshold are considered as subcarriers with damaged signals, and vice versa. Global prior features of the frequency-domain received signal are obtained based on the OFDM signal features of all subcarriers in the frequency-domain received signal. The method for semantically decoding and reconstructing the OFDM signal features of each subcarrier based on its attention weight includes: for subcarriers with damaged signals, extracting OFDM signal features of adjacent subcarriers that are not damaged signals using a multi-scale feature fusion mechanism; fusing and superimposing the extracted OFDM signal features of adjacent subcarriers that are not damaged signals with the global prior features using a residual network, and combining this with the nonlinear mapping process of the residual network to reconstruct the OFDM signal features of the damaged subcarriers, thus obtaining reconstructed signal features; and performing semantic decoding based on the reconstructed signal features of the damaged subcarriers and the OFDM signal features of the undamaged subcarriers to obtain the reconstructed original image.

[0018] In some embodiments of the present invention, the channel state information and the frequency domain received signal are both complex matrices containing amplitude and phase. In the preprocessing module, the method of preprocessing and concatenating the estimated channel state information and the frequency domain received signal includes: extracting the real and imaginary parts of the frequency domain received signal and extracting the real and imaginary parts of the channel state information; concatenating the real part of the frequency domain received signal with the real part of the estimated channel state information in terms of amplitude and phase, and concatenating the imaginary part of the frequency domain received signal with the imaginary part of the estimated channel state information, to obtain joint input data.

[0019] In some embodiments of the present invention, a trained hybrid channel estimator and decoder are employed in the receiving device. The training method of the hybrid channel estimator and decoder includes: acquiring a training set, which includes multiple samples, each sample including a sample image and real channel state information; training the hybrid channel estimator and decoder using the training set, including: encoding the sample image and simulating channel distortion to generate a time-domain received signal carrying image information received by the receiving device; processing the time-domain received signal through an OFDM demodulator to obtain a frequency-domain received signal; obtaining channel state estimation information of the sample based on the frequency-domain received signal by the hybrid channel estimator; obtaining a reconstructed sample image by the decoder based on the channel state estimation information and the frequency-domain received signal; and updating the parameters of the hybrid channel estimator and decoder with the optimization objective of minimizing the difference between the sample image and the reconstructed sample image, as well as the difference between the real channel state information and the channel state estimation information.

[0020] In some embodiments of the present invention, the hybrid channel estimator includes: a pilot preprocessing and embedding module, used to map the pilot signal to a high-dimensional feature space using a linear layer to obtain latent space features, and to embed the latent space features using an embedding layer to obtain pilot embedding features; a dual-branch feature extraction module, used to obtain pilot features containing position and embedding features after subcarrier position encoding of the pilot embedding features, to process the pilot features using a Transformer network to obtain a first feature vector of the pilot signal in the frequency domain, and to process the pilot embedding features using a bidirectional GRU network to obtain a second feature vector of the pilot signal in the time domain; and a channel estimation module, used to fuse the first feature vector and the second feature vector to obtain dynamic fused features, and to process the dynamic fused features using a multilayer perceptron to obtain estimated channel state information.

[0021] In some embodiments of the present invention, the pilot embedding features are encoded by a dual-branch feature extraction module. The encoding method includes: calculating the absolute position vector of each pilot signal using trigonometric functions of different frequencies based on the subcarrier index of each pilot signal in the frequency domain received signal; and adding the absolute position vector of each pilot signal to the pilot embedding features to obtain the pilot features.

[0022] In some embodiments of the present invention, the bidirectional GRU network includes a forward GRU network and a backward GRU network. The bidirectional GRU network processes the pilot embedding features in the following manner: the forward GRU network captures the influence of the previous pilot signal on the subsequent pilot signal from the pilot embedding features in a forward-to-back time sequence to obtain a forward hidden state; the backward GRU network captures the context information of the subsequent pilot signal on the previous pilot signal from the pilot embedding features in a backward-to-front time sequence to obtain a backward hidden state; and the forward hidden state and the backward hidden state are concatenated to obtain a second feature vector of the pilot signal in the time domain.

[0023] According to a second aspect of the present invention, a semantic communication device for OFDM systems is provided for image transmission in high-speed communication scenarios, including unmanned vehicle communication scenarios, drone communication scenarios, and industrial internet scenarios. The device includes: a transmitter for acquiring an original image and converting the original image into a time-domain received signal carrying image information before transmitting it; and a receiver as described in the first aspect of the present invention for receiving the time-domain received signal and reconstructing the original image based on the time-domain received signal.

[0024] Compared with the prior art, the advantages of the present invention are as follows:

[0025] The hybrid channel estimator of the receiving device of this invention extracts the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the frequency domain received signal. Based on the dynamic fusion characteristics, the channel estimation accuracy is effectively improved. The decoder performs semantic decoding and reconstruction based on the channel state information and the frequency domain received signal to obtain the reconstructed original image. This makes the performance of the receiving device no longer passively limited by the accuracy of the channel estimation. Even if there is a certain error in the channel estimation, the implicit equalization strategy is realized by decoding and reconstructing the signal by combining the channel state information and the frequency domain received signal. This avoids amplifying the channel estimation error by using the existing explicit equalization method and improves the performance of source decoding or semantic image reconstruction. Attached Figure Description

[0026] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0027] Figure 1 This is a schematic diagram illustrating the structural principle of a receiving device according to an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating the structural principle of a hybrid channel estimator according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram illustrating the structure and execution principle of the decoder according to an embodiment of the present invention;

[0030] Figure 4This is a schematic diagram illustrating the results of comparing MS-SSIM metrics under two communication system configurations according to an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram showing the results of comparing PSNR indicators under two communication system configurations according to an embodiment of the present invention.

[0032] Figure 6 This is a schematic diagram showing the comparison results of the hybrid channel estimator and baseline model of the present invention under different signal-to-noise ratios according to an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0034] As mentioned in the background section, when the existing technology is applied to high-speed mobile scenarios, the channel estimation accuracy is severely reduced under high dynamic conditions, which further leads to the subsequent explicit equalization amplification of errors and propagation into the recovered symbols, thereby causing the final source decoding or semantic reconstruction performance to deteriorate sharply.

[0035] To address the aforementioned problems, the inventors propose a receiving device for joint decoding and implicit equalization in a high-speed mobile OFDM system. This receiving device acquires a time-domain received signal carrying image information and reconstructs the original image based on the time-domain received signal. The receiving device includes an OFDM demodulator, a hybrid channel estimator, and a decoder. The OFDM demodulator performs OFDM demodulation on the received time-domain received signal to obtain a frequency-domain received signal. The hybrid channel estimator extracts the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the frequency-domain received signal, and estimates the channel state information based on the dynamic fusion characteristics, improving the channel estimation accuracy. The decoder performs semantic decoding and reconstruction based on the channel state information and the frequency-domain received signal to obtain the reconstructed original image. This allows the performance of the receiving device to no longer be passively limited by the accuracy of the channel estimation. Even if there is a certain error in the channel estimation, the implicit equalization strategy is implemented through decoding and reconstruction using the joint channel state information and the frequency-domain received signal, thereby avoiding the amplification of channel estimation errors by using existing explicit equalization methods and improving the performance of source decoding or semantic image reconstruction. In summary, the receiving device of the present invention can effectively resist channel estimation errors in harsh environments with high-speed movement and low SNR, demonstrating excellent image reconstruction robustness.

[0036] According to one embodiment of the present invention, see Figure 1This is a schematic diagram of the structural principle of the receiving device. It includes an OFDM demodulator, a hybrid channel estimator, and a decoder. To better understand the present invention, the various components of the receiving device are described in detail below with reference to specific embodiments.

[0037] I. OFDM Demodulator

[0038] According to one embodiment of the present invention, an OFDM demodulator is used to perform OFDM demodulation on a time-domain received signal received by a receiving device to obtain a frequency-domain received signal. Performing OFDM demodulation on the time-domain received signal includes removing the cyclic prefix (CP) and performing a Fast Fourier Transform (FFT) to obtain the frequency-domain received signal. .

[0039] II. Hybrid Channel Estimator

[0040] According to one embodiment of the present invention, the hybrid channel estimator uses a neural network to extract the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the frequency domain received signal, and estimates the channel state information based on the dynamic fusion characteristics.

[0041] The inventors, through research, recognized that the time-frequency characteristics of highly dynamic channels exhibit decoupling features: in the frequency domain (across subcarriers), the channel response has global correlation, meaning that the fading of different subcarriers is not completely independent but is jointly affected by multipath propagation and frequency-selective fading; in the time domain (across OFDM symbols), channel evolution exhibits local dynamics, meaning that the channel state at the current moment is highly correlated with the channel state at adjacent moments, and this correlation can be described by a Doppler frequency shift and time correlation function. Inspired by this physical characteristic, according to an embodiment of the present invention, a hybrid channel estimator (T-GRU estimator) is designed, the implementation logic of which is as follows:

[0042] First, a Transformer network branch is set up. Taking advantage of the global receptive field of its self-attention mechanism, the global dependencies in the frequency domain dimension are captured from the sparse pilot signal to obtain the first feature vector of the pilot signal in the frequency domain, thus overcoming the limitation of the local receptive field of CNN.

[0043] Then, a bidirectional GRU network branch is set up in parallel with the Transformer network branch. The bidirectional GRU's ability to model time-series data is used to accurately track the smooth evolution trend of the channel in the time domain, and obtain the second feature vector of the pilot signal in the time domain, thus avoiding the gradient vanishing problem of RNN.

[0044] Finally, the outputs of the two branches are fused through a channel estimation module that employs an adaptive fusion mechanism (such as attention-based weighted summation or gating units). This fusion mechanism can dynamically adjust the weights of the two branches according to the current channel conditions. In addition, the fusion result is processed by a multilayer perceptron to obtain the estimated channel state information.

[0045] The above structure enables high-precision recovery of complex time-frequency dynamic characteristics in sparse pilot applications. In other words, the implementation logic of this estimator allows the present invention to obtain the most accurate channel state information estimation possible under highly dynamic, sparse pilot conditions.

[0046] According to one embodiment of the present invention, see Figure 2 This diagram illustrates the structural principle of a hybrid channel estimator. The hybrid channel estimator includes a pilot preprocessing and embedding module, a dual-branch feature extraction module, and a channel estimation module. The three modules are explained below:

[0047] 1) Pilot preprocessing and embedding module

[0048] According to one embodiment of the present invention, the module uses a linear layer to map the pilot signal to a high-dimensional feature space to obtain latent space features, and uses an embedding layer to embed the latent space features to obtain pilot embedding features.

[0049] 2) Dual-branch feature extraction module

[0050] According to one embodiment of the present invention, the module is used to obtain pilot features containing position and embedding features after subcarrier position encoding of pilot embedding features, process the pilot features using a Transformer network to obtain a first feature vector of the pilot signal in the frequency domain, and process the pilot embedding features using a bidirectional GRU network to obtain a second feature vector of the pilot signal in the time domain.

[0051] According to one embodiment of the present invention, subcarrier position encoding is performed on the pilot embedding features through a dual-branch feature extraction module. The encoding method includes: calculating the absolute position vector of each pilot signal using trigonometric functions of different frequencies based on the subcarrier index of each pilot signal in the received signal in the frequency domain; and adding the absolute position vector of each pilot signal to the pilot embedding features to obtain the pilot features. This embodiment achieves at least the following beneficial technical effects: by adding the absolute position vector to the pilot embedding features, the permutation invariance of the subsequent use of a Transformer network with a self-attention mechanism is broken, enabling the Transformer network to perceive the absolute position of different pilots in the frequency band and extract the first feature vector of the obtained pilot signal in the frequency domain, thus achieving more accurate recovery of the fading characteristics of the entire frequency band.

[0052] According to one embodiment of the present invention, the Transformer network includes a multi-head self-attention mechanism and a feedforward neural network layer. The self-attention mechanism is used to capture the global dependency of the pilot signal in the frequency domain, and the feedforward neural network layer outputs the first feature vector of the pilot signal in the frequency domain. The self-attention mechanism works by calculating the query-key inner product between subcarriers of different pilot signals, directly assigning attention weights to pilots at any two positions in the frequency band. This means that even if two subcarriers are far apart in the frequency domain, the network can directly establish a connection between them (i.e., global dependency) in a single layer of computation, thereby comprehensively reconstructing the characteristics of the pilot signals in the entire frequency domain.

[0053] According to an embodiment of the present invention, a bidirectional GRU network includes a forward GRU network and a backward GRU network. The bidirectional GRU network processes the pilot embedding features by: capturing the influence of the preceding pilot signal on the following pilot signal from the pilot embedding features in a forward-to-back time sequence using the forward GRU network, thus obtaining a forward hidden state; capturing the context information of the following pilot signal on the preceding pilot signal from the pilot embedding features in a backward-to-front time sequence using the backward GRU network, thus obtaining a backward hidden state; and concatenating the forward and backward hidden states to obtain a second feature vector of the pilot signal in the time domain. This embodiment achieves at least the following beneficial technical effects: due to the large Doppler frequency shift and short channel coherence time in high-speed mobile scenarios, the channel changes rapidly but continuously between adjacent OFDM pilot symbols. Traditional interpolation methods (such as linear and spline interpolation) struggle to accurately track this non-stationary change. This invention's bidirectional GRU, by forward-preserving the historical temporal state of pilot signals and backward-predicting future temporal states, accurately fits the continuous, gradual physical process of channel state changes between adjacent OFDM pilot symbols based on Doppler frequency shift in high-speed mobile scenarios. It learns this dynamic evolution pattern, and even with sparse pilots, it can supplement intermediate channel states through temporal correlation, thereby capturing the smooth evolution trend of pilots in the time domain and outputting features. This is more suitable for non-stationary, highly dynamic scenarios than fixed interpolation methods.

[0054] 3) Channel estimation module

[0055] According to one embodiment of the present invention, the module is used to fuse a first feature vector and a second feature vector to obtain a dynamic fused feature, and to process the dynamic fused feature through a multilayer perceptron to obtain estimated channel state information.

[0056] According to one embodiment of the present invention, the following adaptive fusion mechanism is used to obtain dynamic fusion features:

[0057] ,

[0058] in, Indicates dynamic fusion characteristics, Represents the weight parameters. Represents the first eigenvector. Represents the second feature vector, weight parameters It is dynamically generated by a lightweight multilayer perceptron (MLP) network based on the current channel conditions, i.e. , This represents the normalized activation function.

[0059] According to one embodiment of the present invention, the fused features The data is fed into a multilayer perceptron (MLP) network for linear mapping, restoring the hidden layer dimension to the physical dimension of the channel matrix, and outputting the real and imaginary parts of the channel response. The real and imaginary parts are then concatenated to finally output the channel state information estimated for the entire two-dimensional time-frequency grid. .

[0060] The technical solution of the above-described hybrid channel estimator embodiment can achieve at least the following beneficial technical effects: it can simultaneously extract complex dynamic characteristics in both the time and frequency domains from sparse pilots, providing high-precision channel prior information for subsequent joint decoding and implicit equalization.

[0061] III. Decoder

[0062] According to one embodiment of the present invention, the decoder utilizes a neural network to perform semantic decoding and reconstruction based on the channel state information and the frequency domain received signal to obtain the reconstructed original image.

[0063] Having obtained channel state information with relatively high accuracy Subsequently, the inventors considered how to break free from the constraints of traditional cascaded architectures. Instead of employing explicit equalization (such as ZF or MMSE) in traditional discrete receivers, they designed a decoder using a joint decoding and implicit equalization strategy, allowing the decoder to autonomously learn how to handle channel distortion during data reconstruction. According to one embodiment of the invention, the joint decoding and implicit equalization strategy does not rely on estimated channel state information. Instead of performing traditional explicit equalization on the received signal, the receiver combines the raw, unequalized received signal (i.e., the demodulated frequency domain received signal) containing inter-carrier interference (ICI), noise, and channel fading with channel state information estimated from the hybrid channel estimator. Together, they serve as joint input data and are fed into the decoder for semantic decoding and reconstruction.

[0064] According to one embodiment of the present invention, see Figure 3This is a schematic diagram illustrating the structure and execution principle of the decoder. Figure 3 In the decoder, the following preprocessing module and feature extraction module are included:

[0065] 4) Preprocessing module

[0066] According to one embodiment of the present invention, the module is used to preprocess the estimated channel state information and the frequency domain received signal and splice them together to obtain joint input data, including channel state data and OFDM signal data corresponding to each subcarrier in the frequency domain received signal.

[0067] According to one embodiment of the present invention, both the estimated channel state information and the frequency domain received signal are complex matrices containing amplitude and phase. In the preprocessing module, the preprocessing and concatenation of the estimated channel state information and the frequency domain received signal includes: extracting the real and imaginary parts of the frequency domain received signal and extracting the real and imaginary parts of the channel state information; concatenating the real part of the frequency domain received signal with the real part of the estimated channel state information in terms of amplitude and phase, and concatenating the imaginary part of the frequency domain received signal with the imaginary part of the estimated channel state information, to obtain joint input data. The joint input data is a four-channel real-valued tensor. The technical solution of this embodiment can achieve at least the following beneficial technical effects: Since neural networks do not directly process complex signals, this invention combines the real and imaginary parts of the frequency domain received signal with... The real and imaginary parts are extracted separately and concatenated along the channel dimension to better utilize neural network processing in subsequent implementations.

[0068] 5) Feature extraction module

[0069] According to one embodiment of the present invention, the feature extraction module utilizes a deep neural network to extract channel state features and OFDM signal features of each subcarrier from the joint input data. Attention weights are assigned to each subcarrier based on these features, and semantic decoding and reconstruction are performed on the OFDM signal features of each subcarrier based on these attention weights to obtain the reconstructed original image. Specifically, a deep neural network based on Swing Transformer or U-Net can be used. Through a multi-layered deep neural network, relevant features of the subcarriers are extracted and fused to obtain channel state features and OFDM signal features. Then, semantic decoding and reconstruction are performed, learning a highly nonlinear mapping process to achieve implicit equalization in the joint decoding and implicit equalization strategy.

[0070] According to one embodiment of the present invention, the joint input data is processed using layers 1 to intermediate layers of a deep neural network to extract the channel state features and OFDM signal features of each subcarrier. Attention weights are calculated in the intermediate layers of the deep neural network (i.e., the layers in the deep neural network that perform feature downsampling and upsampling). Therefore, the OFDM signal features are shallow features extracted by the deep neural network. For example, layers 1 to 5 of the deep neural network are used to process the joint input data to extract the channel state features and OFDM signal features of each subcarrier.

[0071] According to one embodiment of the present invention, in the feature extraction module, the method for assigning attention weights to each subcarrier includes: employing a cross-attention mechanism, using the OFDM signal features of each subcarrier as a query, and the channel state features of each subcarrier as a key and value, calculating the attention weight of each subcarrier and assigning it to each subcarrier; wherein, the inner product between the OFDM signal features and the channel state features corresponding to the subcarrier is positively correlated with the attention weight, and the smaller the attention weight, the worse the signal quality of the subcarrier. This embodiment achieves at least the following beneficial technical effects: In the cross-attention mechanism, the network automatically perceives the channel state quality of the corresponding subcarrier by calculating the inner product of the query and the key, thereby perceiving the signal quality. Therefore, when calculating the attention score, the network assigns a very small attention weight to subcarriers with poor signal quality. This means that the network actively rejects high-noise signals on that subcarrier, thus fundamentally and adaptively suppressing the noise amplification effect.

[0072] According to one embodiment of the present invention, during the process of assigning attention weights to each subcarrier, the deep neural network of the feature extraction module automatically discovers... The network identifies smaller (i.e., deeply fading) subcarrier regions and actively reduces the attention weight for these high-noise subcarriers, instead using the OFDM signal characteristics of surrounding high-quality subcarriers to fill the gaps. This allows the network to dynamically adjust the attention weight for different subcarriers based on channel quality, thereby adaptively suppressing the noise amplification effect at deep fading locations. For example, suppose an image of a car is being transmitted, where the semantic features of the wheels are modulated onto the k-th subcarrier. In a high-speed moving scenario, the k-th subcarrier encounters deep fading (i.e., the channel gain is close to 0). In traditional explicit equalization (such as ZF), the receiver forcibly divides the received signal by the estimated channel state information, causing the background noise at that location to be drastically amplified, completely obscuring the wheel information. However, in the cross-attention mechanism of this invention, the network automatically perceives the extremely poor channel state of the k-th subcarrier by calculating the inner product of Query and Key. Therefore, when calculating the attention weight, the network assigns a very small attention weight to this deeply fading subcarrier, thus fundamentally and adaptively suppressing the noise amplification effect.

[0073] According to one embodiment of the present invention, subcarriers with attention weights less than a preset threshold are considered as subcarriers with damaged signals, and vice versa. Global prior features of the frequency-domain received signal are obtained based on the OFDM signal features of all subcarriers in the frequency-domain received signal. Specifically, the method for semantically decoding and reconstructing the OFDM signal features of each subcarrier based on its attention weight includes: for a subcarrier with damaged signals, extracting OFDM signal features of adjacent subcarriers that are not damaged signals using a multi-scale feature fusion mechanism; fusing and superimposing the extracted OFDM signal features of adjacent subcarriers that are not damaged signals with the global prior features using a residual network, and combining this with the nonlinear mapping process of the residual network to reconstruct the signal features of the damaged subcarriers, thus obtaining reconstructed signal features; and performing semantic decoding based on the reconstructed signal features of the damaged subcarriers and the OFDM signal features of the undamaged subcarriers to obtain the reconstructed original image.

[0074] Schematic, although the attention weight of the k-th subcarrier (e.g., carrying wheel signals) is small (equivalent to suppressing the noise of deep-fading subcarriers), the feature extraction module network can extract OFDM signal features of surrounding high-quality subcarriers (which may carry information related to the vehicle body or chassis) through a multi-scale feature fusion mechanism. At the same time, through residual network connections, the shallowly extracted, incompletely destroyed global prior features (as contextual semantics) and the OFDM signal features of adjacent, undamaged subcarriers are fused and superimposed to assist in reconstructing the signal features of the destroyed subcarriers, thereby enabling the inference and filling of missing wheel information in the high-dimensional feature space.

[0075] The above-described nonlinear mapping process, which does not rely on inverting the channel matrix but instead dynamically isolates signals with poor quality (such as high noise) through cross-attention, and then uses residual network connections and multi-scale feature fusion mechanisms to utilize contextual semantics to complete the missing information, is the essential principle behind the implicit equalization achieved in this invention. Especially for subcarriers where signals are severely damaged due to deep fading or other reasons, the deep neural network in this invention's decoder, through residual network connections and multi-scale feature fusion, utilizes surrounding high-quality features and contextual semantics to learn how to optimally recover or complete missing semantic information under different channel conditions, thereby achieving better reconstruction of the semantic image.

[0076] According to one embodiment of the present invention, a trained hybrid channel estimator and decoder are employed in the receiving device. The training method of the hybrid channel estimator and decoder includes the following steps S1 and S2:

[0077] Step S1: Obtain the training set, which includes multiple samples, each of which includes a sample image and real channel state information.

[0078] According to one embodiment of the present invention, each sample in the training set includes a sample image and corresponding real channel state information, which is a real time-frequency channel state matrix generated by a channel simulation model.

[0079] Step S2: Train the hybrid channel estimator and decoder using the training set.

[0080] According to an embodiment of the present invention, step S2 includes the following steps S21-S23:

[0081] Step S21: After encoding the sample image and simulating channel distortion, a time-domain received signal carrying image information is generated and received by the receiving device.

[0082] According to one embodiment of the present invention, during training forward propagation, the sample image is encoded and converted into an OFDM signal and then transmitted through an emulated channel. The signal is then received by a receiving device, and the emulated receiving device receives the time-domain received signal carrying the image information.

[0083] Step S22: The time-domain received signal is processed by the OFDM demodulator to obtain the frequency-domain received signal. The hybrid channel estimator obtains the channel state estimation information of the sample based on the frequency-domain received signal. The decoder obtains the reconstructed sample image based on the channel state estimation information and the frequency-domain received signal.

[0084] According to one embodiment of the present invention, firstly, the time-domain received signal is demodulated using an OFDM demodulator to generate a frequency-domain received signal containing noise and multipath interference. Secondly, the frequency-domain received signal is processed by a hybrid channel estimator in the same manner as described in the above embodiment to obtain channel state estimation information for the sample. Finally, the reconstructed sample image is obtained by a decoder using the same manner as described in the above embodiment, based on the channel state estimation information and the frequency-domain received signal.

[0085] Step S23: Update the parameters of the hybrid channel estimator and decoder with the optimization objective of minimizing the difference between the sample image and the reconstructed sample image, as well as the difference between the true channel state information and the estimated channel state information.

[0086] According to one embodiment of the present invention, the difference between the sample image and the reconstructed sample image is the reconstruction loss, which can be evaluated based on the pixel-level or perceptual-level difference between the reconstructed sample image and the sample image at the final output of the decoder. For example, the mean squared error (MSE) or multi-scale structural similarity (MS-SSIM) method can be used to calculate the reconstruction loss. The difference between the true channel state information and the estimated channel state information is the channel estimation auxiliary loss, which can be calculated using normalized mean squared error (NMSE). The parameters of the hybrid channel estimator and decoder are updated with the optimization objective of minimizing the weighted sum of the reconstruction loss and the channel estimation auxiliary loss, until a preset number of iterations is reached or the hybrid channel estimator and decoder network converges. At this point, the update stops, resulting in a trained hybrid channel estimator and decoder.

[0087] The technical solution of this embodiment can achieve at least the following beneficial technical effects: by restoring the fidelity of semantic information through the reconstruction loss constraint network, the channel estimation auxiliary loss acts as a physical layer regularizer, guiding the neural network to learn accurate time-frequency fading characteristics, learn meaningful representations, accelerate system convergence, and improve generalization performance. In addition, the application and data processing methods of the hybrid channel estimator and decoder are the same, forcing the decoder to use channel state information as auxiliary information during end-to-end training, implicitly learning how to reverse the reconstruction of signals destroyed in the frequency domain received signal and suppress interference (such as ICI and channel estimation errors). This further avoids the error amplification effect introduced by explicit equalization, and achieves more robust joint decoding and semantic reconstruction in high-speed mobile scenarios. By optimizing through end-to-end training, the reconstruction loss can be minimized directly, rather than the intermediate channel equalization error, thus achieving global optimality. Furthermore, the decoder can compensate for the estimation error by learning prior knowledge (i.e., global prior features as the semantic structure of the image), rather than simply amplifying the channel estimation error. This makes it more robust to channel estimation errors, fundamentally avoiding error propagation effects and making the receiving device more robust to channel estimation errors (CSI estimation error).

[0088] According to an embodiment of the present invention, a semantic communication device for an OFDM system is provided. The device is used for image transmission in high-speed communication scenarios, including unmanned vehicle communication scenarios, drone communication scenarios, and industrial internet scenarios. The device includes: a transmitting end, used to acquire an original image and convert the original image into a time-domain received signal carrying image information before transmitting it; and a receiving device as described in the above embodiment, used to receive the time-domain received signal and reconstruct the original image based on the time-domain received signal.

[0089] To verify the beneficial effects of the present invention, the inventors conducted the following comparative experiments:

[0090] First, different communication systems were formed by combining different receiving devices and transmitting ends, and the performance of these different communication systems was compared and analyzed. The different communication systems include: Scheme 1) combining the receiving device of this invention with a high-performance transmitting end in the prior art to form a complete communication system; Scheme 2) replacing the hybrid channel estimation module in the receiving device of this invention with a traditional channel estimation method, and then combining it with a high-performance transmitting end in the prior art to form a communication system without the hybrid channel estimation module. The traditional channel estimation method can employ a baseline method based on LS estimation and linear interpolation.

[0091] Secondly, under the above two communication system configurations, a comparison was made between the Multi-Scale Structural Similarity Index (MS-SSIM) based on the human visual system for structural information and the Peak Signal-to-Noise Ratio (PSNR) based on pixel-level error. The following comparative experimental results were obtained:

[0092] See Figure 4 This diagram illustrates the comparison of MS-SSIM metrics under two different communication system configurations. The horizontal axis represents the two different schemes, and the vertical axis represents the MS-SSIM metric. Experimental results show that, compared to Scheme 1, the communication system performance without the hybrid channel estimation module exhibits a significant degradation, with MS-SSIM plummeting to 0.62, representing an absolute performance decrease of up to 0.24.

[0093] See Figure 5 This diagram illustrates the comparison of PSNR metrics under two different communication system configurations. The horizontal axis represents the two different schemes, and the vertical axis represents the PSNR metrics, in dB. Experimental results show that, compared to Scheme 1, the PSNR of the communication system without the hybrid channel estimation module drops to 22.2 dB, representing an absolute performance decrease of up to 6.2 dB.

[0094] The quantitative analysis of the two sets of comparative experiments shows that, under the complete system configuration, the image reconstruction performance reaches the optimal level, demonstrating excellent structural fidelity and perceptual quality. This verifies the irreplaceable nature of the complete receiving device of this invention and indicates that the receiving device proposed in this invention plays a decisive foundational role in the overall communication system. Under complex channel conditions, traditional receivers struggle to provide sufficiently high-precision channel information, severely degrading the input quality of subsequent processing modules. In contrast, the receiving device of this invention, as a fundamental support, significantly improves reconstruction quality.

[0095] To further verify the estimation accuracy advantage of the proposed hybrid channel estimator under high dynamic channels, a comparative experiment was conducted between the proposed channel estimation model (i.e., the proposed hybrid channel estimator) and a traditional baseline model (such as a baseline method based on LS estimation and linear interpolation) under different signal-to-noise ratios (SNR). The normalized mean square error (NMSE) was used as the evaluation index for channel estimation accuracy, and the results are as follows: Figure 6 As shown, it is a schematic diagram illustrating the comparison results of the hybrid channel estimator and the baseline model of the present invention under different signal-to-noise ratios.

[0096] Figure 6 The horizontal axis represents different signal-to-noise ratios (SNR), and the vertical axis represents the NMSE index, in dB. As can be seen from the figure, the hybrid channel estimator of this invention significantly outperforms the traditional baseline method across the entire SNR range (0–30 dB). In low SNR scenarios (in noisy and harsh environments), the hybrid channel estimator can more effectively extract channel features from sparse pilots; in high SNR scenarios (e.g., 30 dB), the hybrid channel estimator maintains its advantage, while the baseline method shows limited improvement. This indicates that traditional methods are limited by their linear interpolation assumptions and cannot fully utilize information under high SNR conditions. In contrast, the deep learning-driven hybrid channel estimator of this invention can continuously approximate the true channel value. These results verify the technical effectiveness of the hybrid channel estimator in simultaneously extracting complex dynamic characteristics in both the time and frequency domains from sparse pilots, providing high-precision channel state priors for subsequent decoder processing.

[0097] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0098] This invention can be a system, method, electronic device, computing device, computer-readable medium, and / or computer program product. A computer program product mainly refers to a software product that implements this solution through a computer program.

[0099] A computer-readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. Computer-readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0100] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A receiving device, characterized in that, During the transmission of the original image, the receiving device acquires a time-domain received signal carrying image information and reconstructs the original image based on the time-domain received signal. The receiving device includes: An OFDM demodulator is used to perform OFDM demodulation on the received time-domain signal to obtain the received frequency-domain signal. The hybrid channel estimator uses a neural network to extract the dynamic fusion characteristics of the pilot signal in the time and frequency domains from the pilot signal of the received signal in the frequency domain, and estimates the channel state information based on the dynamic fusion characteristics. The decoder uses a neural network to perform semantic decoding and reconstruction based on the channel state information and the frequency domain received signal to obtain the reconstructed original image.

2. The receiving device according to claim 1, characterized in that, The decoder includes: The preprocessing module is used to preprocess the estimated channel state information and the frequency domain received signal and splice them together to obtain joint input data, including channel state data and OFDM signal data corresponding to each subcarrier in the frequency domain received signal. The feature extraction module uses a deep neural network to extract the channel state features and OFDM signal features of each subcarrier from the joint input data. It assigns attention weights to each subcarrier based on the channel state features and OFDM signal features of each subcarrier, and performs semantic decoding and reconstruction of the OFDM signal features of each subcarrier based on the attention weights of each subcarrier to obtain the reconstructed original image.

3. The receiving device according to claim 2, characterized in that, In the feature extraction module, the methods for assigning attention weights to each subcarrier include: A cross-attention mechanism is adopted, using the OFDM signal characteristics of each subcarrier as the query and the channel state characteristics of each subcarrier as the key and value, to calculate the attention weight of each subcarrier and assign it to each subcarrier. Among them, the inner product between the OFDM signal features and the channel state features corresponding to the subcarrier is positively correlated with the attention weight.

4. The receiving device according to claim 2, characterized in that, Subcarriers with attention weights less than a preset threshold are considered as subcarriers whose signals are corrupted, and those with attention weights greater than a preset threshold are considered as subcarriers whose signals are not corrupted. The global prior features of the frequency domain received signal are obtained based on the OFDM signal features of all subcarriers of the frequency domain received signal. Among them, the methods for semantic decoding and reconstruction of OFDM signal features of each subcarrier based on the attention weight of each subcarrier include: For subcarriers whose signals are damaged, OFDM signal features of adjacent subcarriers whose signals are not damaged are extracted through a multi-scale feature fusion mechanism; By using a residual network, the OFDM signal features of the subcarriers adjacent to the subcarrier whose signals have not been destroyed are fused and superimposed with the global prior features. Combined with the nonlinear mapping process of the residual network, the OFDM signal features of the subcarrier whose signals have been destroyed are reconstructed to obtain the reconstructed signal features. Semantic decoding is performed based on the reconstructed signal characteristics of the damaged subcarriers and the OFDM signal characteristics of the undamaged subcarriers to obtain the reconstructed original image.

5. The receiving device according to claim 2, characterized in that, Both the channel state information and the frequency domain received signal are complex matrices containing amplitude and phase. In the preprocessing module, the estimated channel state information and the frequency domain received signal are preprocessed and concatenated in the following ways: Extract the real and imaginary parts of the frequency domain received signal and extract the real and imaginary parts of the channel state information; The real part of the frequency domain received signal is concatenated with the real part of the estimated channel state information in terms of amplitude and phase, and the imaginary part of the frequency domain received signal is concatenated with the imaginary part of the estimated channel state information to obtain joint input data.

6. The receiving device according to claim 1, characterized in that, In the receiving device, a trained hybrid channel estimator and decoder are employed. The training methods for the hybrid channel estimator and decoder include: Obtain a training set, which includes multiple samples, each of which includes a sample image and real channel state information; Training a hybrid channel estimator and decoder using the training set includes: After encoding the sample image and simulating channel distortion, a time-domain received signal carrying image information is generated and received by the receiving device. The time-domain received signal is processed by an OFDM demodulator to obtain the frequency-domain received signal. The hybrid channel estimator obtains the channel state estimation information of the sample based on the frequency-domain received signal. The decoder obtains the reconstructed sample image based on the channel state estimation information and the frequency-domain received signal. The parameters of the hybrid channel estimator and decoder are updated with the optimization objective of minimizing the differences between the sample image and the reconstructed sample image, as well as the differences between the true channel state information and the estimated channel state information.

7. The receiving device according to claim 1, characterized in that, The hybrid channel estimator includes: The pilot preprocessing and embedding module is used to map the pilot signal to a high-dimensional feature space using a linear layer to obtain latent space features, and then use an embedding layer to embed the latent space features to obtain pilot embedding features. The dual-branch feature extraction module is used to obtain pilot features containing position and embedding features after subcarrier position encoding of pilot embedding features. The pilot features are then processed by a Transformer network to obtain the first feature vector of the pilot signal in the frequency domain. The pilot embedding features are then processed by a bidirectional GRU network to obtain the second feature vector of the pilot signal in the time domain. The channel estimation module is used to fuse the first feature vector and the second feature vector to obtain dynamic fused features. The dynamic fused features are then processed by a multilayer perceptron to obtain estimated channel state information.

8. The receiving device according to claim 7, characterized in that, The pilot embedding features are encoded using a dual-branch feature extraction module to encode subcarrier positions. The encoding methods include: Based on the subcarrier index of each pilot signal in the frequency domain received signal, the absolute position vector of each pilot signal is calculated using trigonometric functions of different frequencies. The absolute position vector of each pilot signal is then added to the pilot embedding feature to obtain the pilot feature.

9. The receiving device according to claim 7, characterized in that, Bidirectional GRU networks include forward GRU networks and backward GRU networks. The methods by which bidirectional GRU networks process pilot embedding features include: By using a forward GRU network to capture the influence of the previous pilot signal on the next pilot signal from the pilot embedding features in a forward-to-back time sequence, the forward hidden state is obtained. By using a backward GRU network in a backward-to-forward temporal order, the contextual information of the subsequent pilot signal relative to the preceding pilot signal in adjacent pilot signals is captured from the pilot embedding features, thus obtaining the backward hidden state. By concatenating the forward hidden state and the backward hidden state, the second eigenvector of the pilot signal in the time domain is obtained.

10. A semantic communication device for OFDM systems, used for image transmission in high-speed communication scenarios, including unmanned vehicle communication scenarios, drone communication scenarios, and industrial internet scenarios, the device comprising: The transmitting end is used to acquire the original image and convert it into a time-domain received signal carrying image information before transmitting it. The receiving apparatus as described in any one of claims 1-9 is used to receive a time-domain received signal and reconstruct the original image based on the time-domain received signal.

Images