A neural network-based ultra-high code rate demodulation method and system

Abstract

The application provides a neural network-based ultrahigh code rate demodulation method and system, relates to the technical field of electric data processing, and comprises the following steps: firstly, acquiring multi-source parameters and a down-conversion sampling sequence of a receiving device, generating a high-dimensional feature tensor through parallel time-frequency representation processing; then, correcting amplitude and phase distortion by using an adaptive equalization technology, and realizing symbol stream alignment through joint timing carrier synchronization; further, generating soft information by using probability decision and decoding auxiliary processing; and finally, outputting demodulation information through iteration refinement. The application embeds a neural network in a demodulation whole process, breaks through the performance bottleneck of a traditional algorithm under a high dynamic channel through feature mapping, parameterized correction and multi-round consistency verification, and significantly improves the transmission reliability and adaptive capacity of a satellite link.

Description

Technical Field

[0001] This invention relates to the field of electrical data processing technology, and more specifically, to a method and system for demodulating ultra-high code rates based on neural networks. Background Technology

[0002] As satellite communication technology evolves towards higher speeds and higher reliability, satellite-to-ground data transmission systems face severe challenges in handling ultra-high code rate transmissions. Dynamic signal interference, such as Doppler shift and phase noise, exacerbates system complexity. Traditional demodulation methods often employ modular architectures, executing equalization, synchronization, and decoding processes in series. This rigid processing flow leads to low resource utilization and insufficient real-time performance. Especially at high code rates, data redundancy becomes prominent. For example, fixed parameter acquisition and preprocessing introduce unnecessary loads, and the rigidity of the process makes it difficult for algorithms to adapt to rapid channel changes, resulting in error accumulation and performance bottlenecks. Existing technologies rely on manual parameter adjustment based on experience, lack intelligent learning mechanisms, and have low efficiency in coordinating software and hardware decisions. Demodulation accuracy further deteriorates in low signal-to-noise ratio scenarios, necessitating a breakthrough in the traditional framework to improve system robustness.

[0003] Therefore, there is an urgent need for an ultra-high code rate demodulation method and system based on neural networks to solve the above-mentioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for demodulating ultra-high code rates based on neural networks, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] In a first aspect, this application provides a method for ultra-high code rate demodulation based on neural networks, comprising:

[0006] Obtain the frequency reference parameters, sampling rate parameters, quantization accuracy parameters, intermediate frequency signal sampling sequence after down-conversion, and their corresponding timestamp information from the ground signal receiving equipment;

[0007] Parallel time-frequency characterization processing is performed based on the sampling sequence and its corresponding timestamp information. By constructing multi-path parallel feature maps on the high-bandwidth sampling sequence, the original sequence is converted into a multi-scale time-series spectral feature tensor. The frequency reference parameters and quantization accuracy parameters are embedded into the feature maps to obtain a unified high-dimensional time-frequency feature tensor.

[0008] The high-dimensional time-frequency feature tensor is adaptively equalized, and the amplitude and phase distortions are corrected layer by layer by constructing a parameterized mapping model. A coarse-to-fine correction process is used to compensate for the rapidly changing channel, resulting in a preliminary equalized symbol representation with probabilistic characteristics.

[0009] Based on the preliminary equalized symbol representation, joint timing and carrier correction processing is performed. By simultaneously estimating the symbol timing offset, phase and frequency deviation in the feature space, and feeding the estimation results back to the feature representation process, a timing-aligned and phase-corrected symbol stream is obtained.

[0010] Based on the symbol stream, probabilistic decision-making and decoding auxiliary feature generation processes are performed. By mapping the symbol stream to a soft information vector and fusing coding constraints and real-time statistical features, a soft decision symbol set is obtained.

[0011] The soft-decision symbol set is iteratively refined by updating the uncertain region during feature reconstruction and performing multiple rounds of consistency verification by combining LDPC codes and real-time statistical features to obtain the demodulated information.

[0012] Secondly, this application also provides an ultra-high code rate demodulation system based on a neural network, comprising:

[0013] The acquisition unit is used to acquire the frequency reference parameters, sampling rate parameters, quantization accuracy parameters, intermediate frequency signal sampling sequence after down-conversion, and their corresponding timestamp information of the ground signal receiving equipment.

[0014] The mapping unit is used to perform parallel time-frequency characterization processing based on the sampling sequence and its corresponding timestamp information. By constructing multi-path parallel feature mapping on the high-bandwidth sampling sequence, the original sequence is converted into a multi-scale time-series spectral feature tensor. The frequency reference parameter and quantization accuracy parameter are embedded into the feature mapping to obtain a unified high-dimensional time-frequency feature tensor.

[0015] The correction unit is used to adaptively equalize the high-dimensional time-frequency feature tensor, correct the amplitude and phase distortion layer by layer by constructing a parameterized mapping model, and use a coarse-to-fine correction process to compensate for the rapidly changing channel, so as to obtain a preliminary equalized symbol representation with probabilistic characteristics.

[0016] The estimation unit is used to perform joint timing and carrier correction processing based on the preliminary equalized symbol representation. By simultaneously estimating the symbol timing offset, phase and frequency deviation in the feature space, and feeding the estimation results back to the feature representation process, a timing-aligned and phase-corrected symbol stream is obtained.

[0017] The processing unit is used to perform probabilistic decision and decoding auxiliary feature generation processing based on the symbol stream, and obtain a soft decision symbol set by mapping the symbol stream into a soft information vector and fusing coding constraints and real-time statistical features.

[0018] The demodulation unit is used to iteratively refine the soft-decision symbol set. It updates the uncertain region during the feature reconstruction process and performs multiple rounds of consistency checks by combining LDPC codes and real-time statistical features to obtain the demodulated information.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention achieves three core technological innovations by constructing an end-to-end intelligent demodulation framework, significantly improving system performance in ultra-high code rate scenarios. First, the parallel time-frequency representation processing based on multi-source parameter fusion breaks through the limitations of traditional modular architectures. By embedding device-intrinsic parameters such as frequency reference and quantization accuracy into feature tensors, it effectively overcomes the resource waste problem caused by data redundancy, enabling the system to maintain stable representation capabilities within the dynamic range of symbol rate (1Msps to 600Msps).

[0021] Secondly, the closed-loop optimization mechanism of adaptive equalization and joint correction solves the error accumulation problem in traditional serial processing by driving amplitude and phase distortion correction and time-frequency deviation through neural network-driven collaborative estimation.

[0022] The integrated design of probabilistic decision-making and iterative refinement in this invention significantly improves the system's fault tolerance to sudden interference through the fusion of hardware and software information and an uncertainty focusing mechanism. Compared to traditional fixed-threshold decision-making methods, this invention utilizes encoding constraint-driven reconstruction optimization to give the demodulation output adaptive correction characteristics, effectively addressing deep fading scenarios in satellite links. The entire technical solution achieves a synergistic improvement in processing efficiency and robustness through process reconstruction and algorithm integration.

[0023] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the ultra-high code rate demodulation method based on neural networks described in this embodiment of the invention;

[0026] Figure 2This is a schematic diagram of the ultra-high code rate demodulation system based on neural networks described in an embodiment of the present invention.

[0027] In the figure: 701, acquisition unit; 702, mapping unit; 703, correction unit; 704, estimation unit; 705, processing unit; 706, demodulation unit. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0030] Example 1:

[0031] This embodiment provides a method for demodulating ultra-high code rates based on neural networks.

[0032] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, S5 and S6.

[0033] Step S1: Obtain the frequency reference parameters, quantization accuracy parameters, and the sampling sequence of the intermediate frequency signal after down-conversion and its corresponding timestamp information from the ground signal receiving equipment;

[0034] It is understandable that the frequency reference parameters in this step typically refer to the nominal frequency and stability indicators of the receiver's local oscillator. These parameters are directly output through the internal crystal oscillator or atomic clock of the device to ensure carrier synchronization accuracy during demodulation. The sampling rate parameter involves the sampling frequency setting of the analog-to-digital converter, read from the hardware configuration register. It determines the time resolution after signal digitization to meet the bandwidth requirements at high code rates. The quantization accuracy parameter represents the number of bits and linearity of the analog-to-digital conversion, obtained from the device datasheet or calibration records, and affects the degree to which the signal's dynamic range is preserved. The intermediate frequency (IF) signal sampling sequence is the baseband or IF digital sample processed by the downconverter, acquired in real time by a high-speed ADC chip. Its sequence length needs to cover multiple symbol periods to capture complete signal characteristics. The corresponding timestamp information is generated by a hardware clock synchronized with GPS or a network time protocol, marking the precise time of each sampling point to compensate for transmission delays and timing jitter in the satellite link. The coordinated acquisition of these data ensures that, in high-speed satellite communication scenarios, the demodulation system can adaptively process based on the actual device status and channel conditions, laying the foundation for subsequent feature extraction and correction.

[0035] Step S2: Perform parallel time-frequency characterization processing based on the sampling sequence and its corresponding timestamp information. By constructing multi-path parallel feature maps on the high-bandwidth sampling sequence, the original sequence is converted into a multi-scale time-series spectral feature tensor. The frequency reference parameters and quantization accuracy parameters are embedded into the feature maps to obtain a unified high-dimensional time-frequency feature tensor.

[0036] Understandably, this step involves constructing multi-path parallel feature maps, utilizing convolutional neural networks with sliding kernels of different sizes to extract short-term time-frequency local patterns, and fusing timestamp information to capture signal temporal dependencies, thereby generating a multi-scale time-series spectral feature tensor. Simultaneously, a frequency reference parameter is embedded as a bias term in the network layer to ensure the consistency of carrier synchronization references, while the quantization accuracy parameter is used to adjust the feature amplitude range, preserving the signal's dynamic characteristics. This design avoids the information loss inherent in traditional single-path processing, enhances adaptability to high code rate signals through parallelization and multi-scale analysis, and enables the feature tensor to effectively encode the spatiotemporal variation patterns of the signal, providing robust input for subsequent equalization and correction. In this step, step S2 includes steps S21, S22, and S23.

[0037] Step S21: Perform time-series dynamic compensation processing on the sampled sequence and its corresponding timestamp information. Analyze the time pattern of the timestamp information through a recurrent neural network, estimate the symbol interval jitter caused by the Doppler effect, and use an interpolation algorithm to resample the sampled sequence to obtain a time-aligned standardized sampled sequence.

[0038] Understandably, this step first analyzes the temporal pattern of the timestamp sequence using a recurrent neural network (RNN) to estimate the symbol interval jitter caused by the Doppler effect. This RNN utilizes its recurrent connection structure to capture long-term temporal dependencies and identify irregular fluctuations in the symbol boundaries caused by relative motion. Subsequently, an interpolation algorithm is used to resample the original sampled sequence, adjusting the sampling point positions based on the estimated time deviation to achieve temporal alignment and generate a standardized sampled sequence. This processing is particularly crucial in high-speed mobile scenarios such as satellite links, where Doppler shift introduces dynamic changes in the symbol interval. Through adaptive learning of the neural network and local correction using interpolation techniques, temporal errors are effectively compensated, providing a stable input foundation for subsequent feature extraction.

[0039] It is understandable that the formulas for symbol interval jitter estimation and resampling in this step are as follows:

[0040]

[0041]

[0042] in, For the estimated symbol interval jitter, For the nominal symbol period, For the Sigmoid function, and For linear layer weights and biases, It is a recurrent neural network. For the input sampling sequence, A timestamp sequence, Here is the GRU weight matrix. This is a time-aligned, standardized sampling sequence. The total length of the input sampling sequence. For the first The length of each input sampling sequence The first sampled sequence of the original input sequence Each sample value, For cubic spline basis functions, This is the index of the resampled sequence. In the first The time offset at each sampling point.

[0043] Step S22: Perform multi-scale spectral feature extraction processing on the time-aligned standardized sampling sequence. By using a parallel convolutional neural network with convolutional kernels of different sizes sliding in the time domain, the subtle spectral fluctuations within the symbol period and the spectral evolution patterns over multiple symbol periods are extracted respectively. The timestamp information is then fused to generate a spatiotemporally correlated feature map, resulting in a preliminary time-series spectral feature tensor.

[0044] Understandably, this step involves constructing a parallel convolutional neural network structure, using convolutional kernels of different sizes to slide across the time domain, capturing both the local fine patterns of the short-term spectrum and the macroscopic trends of the medium- to long-term spectrum. Smaller convolutional kernels focus on extracting subtle spectral fluctuations within a symbol period, while larger kernels capture spectral evolution patterns over multiple symbol periods. This multi-scale analysis approach is particularly suitable for the non-stationary signal characteristics commonly found in satellite communications, effectively addressing spectral broadening caused by relative motion. During processing, the system also deeply integrates timestamp information with spectral features, establishing spatiotemporal correlations through feature concatenation and weight adjustment mechanisms. This ensures that the generated feature mapping not only includes spectral amplitude information but also embeds the temporal evolution characteristics of the signal. The resulting preliminary time-series spectral feature tensor forms a structured feature representation, where different dimensions correspond to time, frequency, and multi-scale feature information, providing a rich feature foundation for subsequent signal processing stages.

[0045] The multi-scale convolution feature extraction formula in this step is as follows:

[0046]

[0047]

[0048] in, The output feature of the i-th convolutional layer has a dimension of 64. for function, The kernel size represents the analysis time for short (3 points), medium (5 points), and long (7 points). For the first Each convolutional kernel size, For the time-aligned normalized sampling sequence at position The value at that position. Where l is the starting index of the convolution operation. For the first The weight matrix of each convolutional kernel. The convolution bias vector has a dimension of 64. The concatenated feature vector has a dimension of 193. for, and These represent the output features of the 1st, 2nd, and 3rd convolutional layers, respectively. It is a timestamp scalar.

[0049] Step S23: Perform parameterized feature enhancement processing on the preliminary time-series spectral feature tensor. Embed the frequency reference parameter and quantization accuracy parameter as bias terms into the feature channel through a fully connected neural network, and perform weighted fusion of multi-scale features to obtain a unified high-dimensional time-frequency feature tensor.

[0050] Understandably, this step first performs a nonlinear transformation on each feature channel using a fully connected neural network. A frequency reference parameter is embedded as a bias term in the network layer to adjust the baseline offset of the feature distribution, compensating for systematic deviations caused by local oscillator frequency drift in the satellite link. Simultaneously, a quantization accuracy parameter is introduced as another bias term to control the dynamic range of the feature amplitude, avoiding information loss due to limitations in analog-to-digital conversion accuracy. Subsequently, this step performs weighted fusion of multi-scale features, calculating the contribution weights of features at different scales through an attention mechanism. Components highly correlated with channel dynamics are preferentially retained, such as those sensitive to the Doppler effect from short-term spectral fluctuations, ultimately generating a unified high-dimensional time-frequency feature tensor. This processing, through parameterized embedding, deeply integrates device-specific parameters with learned features, enhancing the adaptability of feature representation to hardware errors and channel variations. It is particularly suitable for high-code-rate scenarios requiring a balance between accuracy and complexity.

[0051] The formulas for parameter embedding and feature weighted fusion in this step are shown below:

[0052]

[0053]

[0054] in, For fully connected output features, For the input feature tensor, The fully connected weight matrix has a size of 193×512. This is a fully connected bias vector with dimension 512. Here is the frequency reference parameter (10 MHz), and 1 represents an all-one vector. For quantization precision (12 bits). To output a high-dimensional time-frequency feature tensor, For element-wise multiplication, The row vector representing the attention weight vector.

[0055] Step S3: Adaptively equalize the high-dimensional time-frequency feature tensor, construct a parameterized mapping model to correct the amplitude and phase distortion layer by layer, and use a coarse-to-fine correction process to compensate for the rapidly changing channel, so as to obtain a preliminary equalized symbol representation with probabilistic features.

[0056] It is understandable that this step constructs a multi-layer neural network structure through a parameterized mapping model, progressively correcting nonlinear distortion in the feature tensor: first, a convolutional layer is used for coarse correction to quickly capture the global distortion pattern; then, a recurrent layer is used for fine correction to precisely adjust the local phase deviation, forming a pipeline from coarse to fine. In this step, step S3 includes steps S31, S32, and S33.

[0057] Step S31: Perform coarse equalization correction on the high-dimensional time-frequency feature tensor. Perform convolution operation on the feature space through a parallel convolutional neural network to extract local fine features and capture global features, and compensate for amplitude distortion to obtain a preliminary corrected feature tensor;

[0058] Understandably, this step first performs convolution operations in the feature space by sliding multi-sized convolutional kernels through a convolutional neural network architecture. Smaller kernels focus on extracting local fine features, such as amplitude fluctuations at symbol boundaries, while larger kernels capture global features, such as macroscopic distortion patterns caused by channel attenuation. This multi-scale analysis effectively addresses amplitude distortion caused by multipath effects and noise superposition in satellite links. The neural network learns weights to adaptively correct distortion components, generating a preliminary corrected feature tensor. This step, as the initial stage of the equalization process, provides the foundation for subsequent fine equalization. Its coarse correction characteristics ensure rapid signal stabilization under rapidly changing channels, preventing error propagation, while preserving the structural integrity of the signal through spatial feature extraction.

[0059] The formula for multi-scale convolution amplitude equalization is shown below:

[0060]

[0061]

[0062] in, Multi-scale features extracted by convolutional neural networks For the input high-dimensional time-frequency feature tensor, For the first Each convolutional kernel weight matrix has a size of [size missing]. ×128, For the first There are 128 convolutional bias vectors. This is a concatenation of all convolutional outputs. To output the preliminary corrected feature tensor, the size ×384.

[0063] Step S32: The preliminary correction feature tensor is refined and balanced. The continuous phase error patterns in the preliminary correction feature tensor are learned by a recurrent neural network, and the phase distortion is corrected point by point. The weights are dynamically adjusted in combination with the rapidly changing channel characteristics of the satellite link to obtain the refined balanced features.

[0064] It is understandable that in this step, the recurrent neural network, through its temporal memory characteristics, learns the continuous phase error patterns in the feature tensor. Specifically, it captures the long-term dependencies of phase changes through a gating mechanism (LSTM unit) and analyzes the regularity of phase jumps between symbols point by point. During processing, the network dynamically adjusts the weight allocation of the hidden state according to the rapidly changing channel characteristics unique to satellite links (such as periodic frequency offsets caused by the high-speed motion of low-Earth orbit satellites), making the network more sensitive to sudden phase jitter. This dynamic weight adjustment mechanism adaptively optimizes network parameters by calculating the channel change gradient in real time, ensuring accurate compensation for phase distortion at the symbol level. The final refined equalized features retain the amplitude correction results and significantly improve phase continuity, providing a high-quality feature base for subsequent symbol decision. The innovation of this processing step lies in combining temporal modeling with dynamic weight adjustment, effectively addressing the non-stationary characteristics of phase errors in highly dynamic satellite environments.

[0065] The LSTM phase error correction formula is shown below:

[0066]

[0067]

[0068] in, The hidden states and cell states are represented by LSTM, with a dimension of 256. For input feature number One time step, 384 dimensions For the first The hidden state and cell state at each time step. Here is the LSTM weight matrix. This is the bias vector of the LSTM. To output refined and balanced features, and These are the attention weight matrix and the bias, respectively. It is a channel state vector with 16 dimensions, containing real-time estimated parameters.

[0069] Step S33: Perform probabilistic symbol generation processing on the refined equilibrium features. Output the probability distribution of each symbol through a fully connected neural network and fuse uncertainty estimates to obtain a preliminary equilibrium symbol representation with probabilistic features.

[0070] Understandably, this step first uses a fully connected neural network to nonlinearly transform the feature vector of each symbol, mapping the high-dimensional features to a probability distribution. The network output layer uses a Softmax activation function to generate probability values ​​for each symbol corresponding to different modulation constellation points, thus quantifying the uncertainty of symbol decision. Simultaneously, the system integrates uncertainty estimation, such as through Monte Carlo Dropout or confidence calculation, to evaluate the reliability of the probability output and avoid misjudgments under low signal-to-noise ratio. Finally, a preliminary balanced symbol representation with probabilistic features is obtained, whose soft information characteristics provide rich input for subsequent decoding. This step preserves uncertainty information through probabilistic output, enhancing the system's fault tolerance to sudden errors and improving the overall robustness of demodulation.

[0071] The formula for generating the Softmax probability distribution is shown below:

[0072]

[0073]

[0074] in, For symbols Belongs to constellation points The probability, The Softmax temperature parameter has a value of 2. For the output of a fully connected layer, dimension , The numbers of the constellation points. To output the probability distribution, The probability distribution of each symbol belonging to different modulation constellation points. Number of samples taken in Monte Carlo.

[0075] Step S4: Perform joint timing and carrier correction processing based on the preliminary equalized symbol representation. By simultaneously estimating the symbol timing offset, phase and frequency deviation in the feature space, and feeding the estimation results back to the feature representation process, a timing-aligned and phase-corrected symbol stream is obtained.

[0076] Understandably, this step constructs a dual-branch neural network structure. The timing estimation branch uses a convolutional neural network to detect symbol boundary jitter, while the carrier correction branch utilizes a recurrent neural network to learn frequency offset timing features. The outputs of the two branches are weighted and fused through an attention mechanism to form a joint correction vector. The estimation result is then fed back to the feature representation layer for parameter fine-tuning, forming a closed-loop optimization mechanism. This joint processing method effectively overcomes the limitations of the mutual constraints between timing and carrier recovery in traditional serial processing, and is particularly suitable for scenarios with timing jitter and rapid frequency changes in high-dynamic satellite links. In this step, step S4 includes steps S41, S42, and S43.

[0077] Step S41: Perform symbol timing offset estimation processing on the preliminary equalized symbol representation. Use a parallel convolutional neural network to slide multi-size convolutional kernels on the symbol sequence to capture subtle timing jitter and identify macro-temporal drift. Detect temporal irregularities at symbol boundaries. Utilize the adaptive learning capability of the convolutional neural network to extract timing error features. Use an interpolation algorithm to resample the detected offset to obtain a preliminary timing-corrected symbol sequence.

[0078] Understandably, this step first uses a convolutional neural network to slide multi-sized convolutional kernels across the symbol sequence to detect temporal irregularities at symbol boundaries. Smaller kernels capture subtle timing jitter, while larger kernels identify macroscopic temporal drift. The adaptive learning capability of the neural network is used to extract timing error features. Subsequently, the system uses interpolation algorithms to resample the detected offsets, such as adjusting the symbol interval through cubic splines or linear interpolation, to achieve timing alignment and generate a preliminary timing-corrected symbol sequence. This step effectively improves the robustness and accuracy of timing estimation through local feature learning of the CNN and smoothing correction by the interpolation algorithm, providing a stable symbol stream foundation for subsequent carrier recovery.

[0079] The formula for the preliminary timing correction symbol sequence is as follows:

[0080]

[0081]

[0082] in, For the estimated timing offset, For symbol period, It is the hyperbolic tangent function. Let be the weight matrix. This is a vector concatenation operation. Let l be the output feature vectors of the first, second, and third convolutional layers at position l. This is the bias vector for timing correction. For the initial timing correction of the symbol sequence, For sequence length, It is a cubic spline curve. This is the index variable for the new sequence after resampling.

[0083] Step S42: Perform joint estimation of phase and frequency deviation on the preliminary timing correction symbol sequence. Use a recurrent neural network to learn the Doppler spectrum hidden in the symbol stream and dynamically compensate for phase rotation to obtain the frequency deviation corrected symbol stream.

[0084] Understandably, this step utilizes the temporal modeling capabilities of a recurrent neural network (LSTM unit) to learn the implicit Doppler spectral dynamics within the symbol stream. The network continuously tracks frequency offset trends through its hidden states, capturing the continuous phase rotation patterns caused by the relative motion of the satellite. During processing, the network not only learns instantaneous frequency offset values ​​but also memorizes historical frequency offset patterns through a gating mechanism, thereby predicting the phase change trajectory of subsequent symbols. This temporal correlation learning enables the system to adapt to the periodic frequency offset variations unique to low-Earth orbit satellite scenarios, such as the symmetrical Doppler frequency shift curve generated when a satellite passes overhead.

[0085] In its implementation, the network takes a sequence of symbols as input and gradually constructs a frequency offset state model through its recursive connection structure. For each symbol time, the network simultaneously outputs phase correction and frequency compensation values, achieving joint estimation. The dynamic compensation mechanism adjusts the weight distribution of the network's hidden states in real time, keeping the system sensitive to sudden frequency offset jumps (such as frequency offset abrupt changes caused by orbital maneuvers). During the compensation process, phase rotation correction uses complex multiplication to accurately correct phase deviations while maintaining the symbol amplitude. This step integrates frequency offset estimation and phase compensation into a single learning task, overcoming the limitation of traditional phase-locked loops requiring step-by-step processing. Through RNN temporal modeling, the system can autonomously learn Doppler dynamic characteristics from data, achieving accurate tracking without pre-setting a satellite orbit model.

[0086] The formulas for LSTM frequency offset tracking and phase rotation compensation are shown below:

[0087]

[0088]

[0089] in, These are the phase correction amount and frequency offset estimates. For output layer weights, For output layer bias, For LSTM hidden states, The output frequency offset correction symbol stream is a complex matrix. For the input symbol sequence, the first A symbol, dimension M, This is the phase correction amount. This is the frequency offset estimate. Pi is the mathematical constant of a circle.

[0090] Step S43: Perform feedback alignment optimization on the frequency offset corrected symbol stream. Calculate the confidence weight of each position in the symbol stream through an attention mechanism, identify the symbol intervals that need to be realigned using the weighted feature map, and dynamically adjust the symbol interval using an interpolation algorithm to obtain a time-aligned and phase-corrected symbol stream. During the adjustment process, the adjustment result is backpropagated to the feature extraction layer through a closed-loop feedback mechanism.

[0091] This step, as we understand it, calculates the confidence weight of each position in the symbol stream using an attention mechanism, focusing on regions with large timing errors (such as segments with significant symbol boundary jitter). It then uses weighted feature maps to identify symbol intervals requiring realignment. Specifically, this step first extracts local temporal features of the symbol stream through convolutional layers, and then uses a self-attention mechanism to calculate the correlation weights between symbols. For low-confidence regions (such as symbols with low signal-to-noise ratios), higher attention weights are assigned, and an interpolation algorithm is used to dynamically adjust the symbol intervals. During the adjustment process, combined with the previously estimated timing error distribution, the adjustment results are backpropagated to the feature extraction layer through a closed-loop feedback mechanism, achieving multi-round iterative optimization. This effectively solves the problem of residual timing errors caused by multipath effects and noise superposition in satellite links.

[0092] The correction formula for the residual error of the attention mechanism is as follows:

[0093]

[0094]

[0095] in, This is the attention weight matrix. The amplitude characteristic matrix, This is the query transformation weight matrix in the self-attention mechanism. This is the key transformation weight matrix in the self-attention mechanism. For symbol period parameter, For the output symbol stream, timing alignment and phase correction are performed. This is a frequency offset correction symbol stream.

[0096] Step S5: Perform probabilistic decision and decoding auxiliary feature generation processing based on the symbol stream. By mapping the symbol stream to a soft information vector and fusing coding constraints and real-time statistical features, a soft decision symbol set is obtained.

[0097] Understandably, this step maps the symbol stream into a soft information vector using a neural network, which encodes the likelihood probability information of each symbol. Subsequently, it fuses coding constraints, such as the parity check matrix relationship of the forward error correction code, and statistical features, such as channel condition estimation, to generate a soft-decision symbol set with a confidence distribution. This processing method preserves uncertainty through soft information, enhancing the fault tolerance of the demodulation system in low signal-to-noise ratio and high interference scenarios. It is particularly suitable for the dynamic channel conditions commonly found in satellite links, providing reliable input for subsequent refinement. In this step, step S5 includes steps S51, S52, and S53.

[0098] Step S51: Perform soft information vector generation processing on the timing-aligned and phase-corrected symbol stream. Convert the feature vector of each symbol in the symbol stream into a probability distribution through a fully connected neural network, and calculate the log-likelihood ratio to obtain the initial soft information vector.

[0099] Understandably, this step first maps the symbol stream using a fully connected neural network. This network consists of multiple hidden layers, converting the feature vector of each symbol into a probability distribution (the Softmax probability distribution generation formula in step S33). The number of output layer nodes corresponds to the possible symbol values ​​of the modulation scheme (the four output nodes of QPSK modulation represent the probabilities of different constellation points), thereby quantifying the likelihood of each symbol belonging to a specific modulation point. Subsequently, the system calculates the log-likelihood ratio, a soft-information metric that characterizes the confidence level of a decision by comparing the log-likelihood ratio of the likelihood values ​​for each bit being 0 or 1, replacing traditional hard decisions to retain more uncertainty information. The resulting initial soft-information vector provides probabilistic input for subsequent decoding stages. In the high code rate scenario of satellite communication, the adaptive learning capability of the fully connected network effectively captures hidden patterns in the symbol stream, and combined with the soft-decision characteristics of the log-likelihood ratio, enhances the system's robustness in low signal-to-noise ratio and rapid channel changes, laying the foundation for error control in the overall demodulation process.

[0100] The formula for calculating the log-likelihood ratio is shown below:

[0101]

[0102] in, for No. The log-likelihood ratio of bits. For bit condition indexing, For symbols Belongs to constellation points The probability of.

[0103] Step S52: Perform encoding constraint fusion processing on the initial soft information vector, and iteratively update the soft information through a graph neural network to fuse the verification relationship of commonly used LDPC codes in satellite links to obtain the enhanced soft information vector;

[0104] Understandably, this step first uses a graph neural network to model the parity check matrix of LDPC (Low-Density Parity-Check) codes commonly used in satellite links as a bipartite graph structure. Variable nodes correspond to encoded bits, and parity nodes represent linear constraint relationships. Utilizing the message passing mechanism of the graph neural network, soft information is iteratively exchanged between variable nodes and parity nodes. In each iteration, nodes update their own state based on the received probability messages, gradually incorporating encoded constraints to correct transmission errors. This step effectively fuses the code's parity relationships, enhancing the reliability of the soft information. In high code rate scenarios of satellite links, the sparsity of LDPC codes enables graph neural networks to efficiently handle large-scale constraints, adapting to low signal-to-noise ratios and sudden interference. Through multiple rounds of iterative optimization, the initial soft information vector gradually converges, yielding an enhanced soft information vector.

[0105] The LDPC Tanner graph message passing formula is shown below:

[0106]

[0107]

[0108] in, For the first Variable nodes in the next iteration To the verification node The news For variable nodes Prior information, For the first Verification nodes in the next iteration To variable node The news For variable nodes A set of connected verification nodes. For the first Verification nodes in the next iteration To the variable node The news To verify the node A set of connected variable nodes. For the first Variable nodes in the next iteration To the verification node The news for The absolute value of.

[0109] Step S53: Perform statistical feature integration processing on the enhanced soft information vector. By dynamically weighting the channel statistical parameters, the weighted statistical features are fused with the enhanced soft information vector, and the output value is normalized to obtain the soft decision symbol set.

[0110] Understandably, this step first dynamically weights channel statistical parameters through an attention mechanism. These parameters include real-time estimated signal-to-noise ratio (SNR), Doppler rate of change, and other indicators reflecting link quality. The attention network calculates the correlation between each soft information component and the current channel state, assigning greater weight to highly correlated features to form a weighted statistical feature vector. Next, the system fuses the weighted statistical features with the enhanced soft information vector, using feature concatenation and fully connected layer transformation to achieve information integration. During this process, the network learns the reliability patterns of soft information under different channel conditions, for example, automatically reducing the weight of error-prone locations during periods of low SNR. Finally, normalization processing keeps the output values ​​within a stable range, generating a soft decision symbol set that conforms to probability axioms. This allows the invention to effectively cope with rapid fluctuations in signal quality, ensuring that the soft decision result retains coding gain while adapting to real-time channel changes.

[0111] The channel statistical weighting formula for the attention mechanism is shown below:

[0112]

[0113]

[0114] in, For the first Attention weights for each symbol The Softmax activation function is used. For the first The enhanced soft information vector of each symbol, To query the weight matrix, Key weight matrix, For the first Channel statistical parameter matrix of symbols, For soft decision symbol set, for, For the first The mean of the soft information weighted vector of each symbol before normalization. For the first The standard deviation of the soft information weighted vector of each symbol before normalization This indicates transpose.

[0115] Step S6: Iteratively refine the soft decision symbol set by updating the uncertain region during feature reconstruction and performing multiple rounds of consistency verification with LDPC codes and real-time statistical features to obtain the demodulated information.

[0116] Understandably, this step constructs a feature reconstruction mechanism, first using an attention network to identify low-confidence symbol regions, and then employing a generative network to reconstruct local features of the target region. During the reconstruction process, the system integrates coding constraints (such as the check relationship of LDPC codes) and real-time statistical features (such as channel state information), continuously optimizing the symbol representation through multiple rounds of forward-feedback loops. Each iteration performs a consistency check to verify the compatibility of the symbol stream with the coding structure, and adjusts the weights and updates the features of symbols that do not meet the check conditions. This progressive refinement method effectively solves the decision ambiguity problem caused by noise accumulation and interference at high code rates, significantly improving the reliability of the demodulation output. In this step, step S6 includes steps S61, S62, and S63.

[0117] Step S61: Perform uncertain region identification processing based on the soft decision symbol set, calculate the confidence weight of each symbol through the attention mechanism and focus on the low reliability region to obtain the position mapping of the symbol to be refined;

[0118] Understandably, this step first dynamically calculates the confidence weight of each symbol using an attention mechanism. Specifically, it employs a self-attention network to analyze the correlation between symbols and quantifies the reliability index based on the log-likelihood ratio versus variance. For low-reliability regions (such as periods of sudden changes in signal-to-noise ratio or symbols with flattened probability distributions due to sudden interference), this step generates a high-weighted focused signal.

[0119] During processing, the attention network constructs a mapping between symbol positions and confidence levels, detecting abnormal fluctuations in the probability distribution through a sliding window. For example, when a satellite link experiences transient deep fading, the probability distribution of affected symbols exhibits multi-peak characteristics, and the attention mechanism marks such regions as high-priority processing targets. The final generated symbol position mapping is essentially a binary mask matrix, where high-weight positions identify the symbol indices that require priority optimization. This step transforms the traditional uniform refinement strategy into a targeted optimization mode. Through the dynamic allocation of attention weights, the system can concentrate limited computing resources on the key regions that have the greatest impact on the overall bit error rate. In the high code rate scenario of satellite communication, this invention improves the efficiency of iterative refinement and avoids redundant processing of high-confidence symbols. Simultaneously, the generation of the position mapping enables subsequent reconstruction processing to employ a non-uniform computing strategy, achieving intelligent scheduling of system resources.

[0120] The formulas for calculating the self-attention confidence weights and generating the location mapping are shown below:

[0121]

[0122]

[0123] in, For the first Attention weights for each symbol For soft decision symbol set, The query weight matrix for the soft-decision symbol set. The key weight matrix of the soft-decision symbol set. For location mapping, This is for variance calculation.

[0124] Step S62: Perform symbol reconstruction processing based on the symbol position mapping to be refined. Use a generative network driven by coding constraints to locally reconstruct low-confidence symbols and optimize the reconstruction parameters by combining satellite link statistical features to obtain the updated symbol representation.

[0125] Understandably, this step uses a generative adversarial network to locally reconstruct the target symbols. Encoding constraints (such as the parity check matrix relationship of LDPC codes) are embedded as prior knowledge in the network loss function, guiding the generator to output a symbol sequence that conforms to the transmission protocol. Simultaneously, it integrates satellite link statistical characteristics, including real-time estimated signal-to-noise ratio and preset Doppler shift rate, and optimizes the weight parameters of the generative network through gradient descent, enabling the reconstruction process to adapt to rapid channel fluctuations. The resulting updated symbol representation significantly improves the decision quality in low-reliability regions, reduces the risk of error propagation, and provides a more robust input basis for subsequent consistency checks, effectively addressing symbol distortion issues caused by deep fading or sudden interference in satellite links.

[0126] The generative adversarial network symbol reconstruction formula is shown below:

[0127]

[0128] in, Hide the generator's state. for function, and For the generator input layer weights and biases, Generate weights for the conditions.

[0129] Step S63: Perform multi-round consistency verification processing based on the updated symbol representation, verify the compatibility between the symbol stream and the encoding structure through iterative feedback loop, and output the final demodulation information that meets the verification conditions.

[0130] Understandably, this step first inputs the symbol stream into a constraint verification network constructed from the LDPC parity-check matrix. A message-passing algorithm is used to detect inconsistencies between the symbol sequence and the parity-check equation round by round. Subsequently, for the locations of symbols that fail verification, a gradient backpropagation mechanism is used to dynamically adjust the symbol probability distribution, and the step size is optimized by combining the real-time statistical characteristics (noise variance) of the satellite link. During processing, the system ensures output reliability through a triple verification mechanism: first, symbol-level verification, using local verification nodes to verify the compatibility of adjacent symbols; second, frame-level verification, using cyclic redundancy check codes to verify the integrity of data frames; and third, system-level verification, using historical demodulation data to establish an error pattern library for cross-validation. Each iteration generates a confidence evaluation index. When the pass rate for three consecutive rounds of verification exceeds a set threshold, the system determines that the symbol stream has reached a stable state, and then generates the final demodulation information.

[0131] The formulas for iterative message passing verification and gradient optimization are shown below:

[0132]

[0133]

[0134] in, For the first During round iteration, verify the message from the node to the variable node. It is the inverse hyperbolic tangent function. For variable nodes The reconstructed log-likelihood ratio, Let be the adjusted log-likelihood ratio for round t. For reconstructing the symbol sequence of the first One element, Let t be the verification loss value for round t.

[0135] Example 2:

[0136] like Figure 2 As shown, this embodiment provides an ultra-high code rate demodulation system based on a neural network. See [link to documentation]. Figure 2 The system includes an acquisition unit 701, a mapping unit 702, a correction unit 703, an estimation unit 704, a processing unit 705, and a demodulation unit 706.

[0137] The acquisition unit 701 is used to acquire the frequency reference parameters, sampling rate parameters, quantization accuracy parameters, intermediate frequency signal sampling sequence after down-conversion, and their corresponding timestamp information of the ground signal receiving equipment.

[0138] The mapping unit 702 is used to perform parallel time-frequency characterization processing based on the sampling sequence and its corresponding timestamp information. By constructing multi-path parallel feature mapping on the high-bandwidth sampling sequence, the original sequence is converted into a multi-scale time-series spectral feature tensor. The frequency reference parameter and quantization accuracy parameter are embedded in the feature mapping to obtain a unified high-dimensional time-frequency feature tensor.

[0139] The correction unit 703 is used to perform adaptive equalization processing on the high-dimensional time-frequency feature tensor, and to perform layer-by-layer correction of amplitude and phase distortion by constructing a parameterized mapping model, and to achieve compensation for the rapidly changing channel by adopting a coarse-to-fine correction process, so as to obtain a preliminary equalized symbol representation with probabilistic characteristics.

[0140] Estimation unit 704 is used to perform joint timing and carrier correction processing based on the preliminary equalization symbol representation. By simultaneously estimating symbol timing offset, phase and frequency deviation in the feature space, and feeding the estimation results back to the feature representation process, a timing-aligned and phase-corrected symbol stream is obtained.

[0141] Processing unit 705 is used to perform probabilistic decision and decoding auxiliary feature generation processing based on the symbol stream, and obtain a soft decision symbol set by mapping the symbol stream into a soft information vector and fusing coding constraints and real-time statistical features;

[0142] The demodulation unit 706 is used to iteratively refine the soft decision symbol set. By updating the uncertain region during the feature reconstruction process and combining the LDPC code with real-time statistical features for multiple rounds of consistency verification, the demodulated information is obtained.

[0143] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0144] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0145] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for demodulating ultra-high code rates based on neural networks, characterized in that, include: Acquire the frequency reference parameters, quantization accuracy parameters, and the sampling sequence of the intermediate frequency signal after down-conversion, along with its corresponding timestamp information, of the ground signal receiving equipment; The sampling sequence and its corresponding timestamp information are subjected to time-series dynamic compensation processing. The time-series pattern of the timestamp information is analyzed by recurrent neural network, the symbol interval jitter caused by the Doppler effect is estimated, and the sampling sequence is resampled by interpolation algorithm to obtain a time-aligned standardized sampling sequence. Multi-scale spectral feature extraction is performed on time-aligned standardized sampling sequences. A parallel convolutional neural network with multi-sized convolutional kernels slides in the time domain to extract subtle spectral fluctuations within a symbol period and capture spectral evolution patterns over multiple symbol periods. Timestamp information is then fused to generate spatiotemporally correlated feature maps, resulting in a preliminary temporal spectral feature tensor. The preliminary time-series spectral feature tensor is subjected to parameterized feature enhancement processing. The frequency reference parameter and quantization accuracy parameter are embedded as bias terms into the feature channel through a fully connected neural network. The multi-scale features are then weighted and fused to obtain a unified high-dimensional time-frequency feature tensor. Coarse equalization correction is performed on the high-dimensional time-frequency feature tensor. Convolution operations are performed in the feature space by sliding multi-size convolution kernels in a parallel convolutional neural network to extract local fine features and capture global features, and to compensate for amplitude distortion, thus obtaining a preliminary corrected feature tensor. The preliminary correction feature tensor is refined and balanced. The continuous phase error patterns in the preliminary correction feature tensor are learned by a recurrent neural network, and the phase distortion is corrected point by point. The weights are dynamically adjusted in combination with the rapidly changing channel characteristics of the satellite link to obtain the refined balanced features. The refined equilibrium features are processed by probabilistic symbol generation. The probability distribution of each symbol is output through a fully connected neural network and the uncertainty estimate is fused to obtain a preliminary equilibrium symbol representation with probabilistic features. The initial equalized symbol representation is processed by symbol timing offset estimation. A parallel convolutional neural network slides multi-size convolutional kernels on the symbol sequence to capture subtle timing jitter and identify macro-temporal drift. Temporal irregularities at symbol boundaries are detected. Timing error features are extracted using the adaptive learning capability of the convolutional neural network. An interpolation algorithm is used to resample the detected offsets to obtain an initial timing-corrected symbol sequence. The initial timing correction symbol sequence is subjected to joint estimation of phase and frequency deviation. The Doppler spectrum implicit in the symbol sequence is dynamically learned and the phase rotation is dynamically compensated through a recurrent neural network to obtain the frequency deviation corrected symbol sequence. The frequency offset corrected symbol sequence undergoes feedback alignment optimization. An attention mechanism is used to calculate the confidence weight at each position in the symbol sequence. The weighted features are then used to identify the symbol intervals requiring realignment, and an interpolation algorithm is employed to dynamically adjust the symbol spacing. A time-aligned and phase-corrected symbol sequence is obtained. During the adjustment process, the adjustment result is backpropagated to the feature extraction layer through a closed-loop feedback mechanism. The timing-aligned and phase-corrected symbol sequence is processed to generate a soft information vector. The feature vector of each symbol in the symbol sequence is converted into a probability distribution through a fully connected neural network, and the log-likelihood ratio is calculated to obtain the initial soft information vector. The initial soft information vector is subjected to encoding constraint fusion processing, and the soft information is iteratively updated through a graph neural network to fuse the verification relationship of the satellite link LDPC code, thereby obtaining the enhanced soft information vector; Channel state information integration processing is performed on the enhanced soft information vector. By dynamically weighting the channel state information, the weighted channel state information is fused with the enhanced soft information vector, and the output value is normalized to obtain the soft decision symbol set. The soft-decision symbol set is iteratively refined by updating the uncertain region during feature reconstruction and performing multiple rounds of consistency verification in combination with the LDPC code check relationship and channel state information to obtain the demodulated information.

2. A high code rate demodulation system based on a neural network, characterized in that, include: The acquisition unit is used to acquire the frequency reference parameters, sampling rate parameters, quantization accuracy parameters, intermediate frequency signal sampling sequence after down-conversion, and their corresponding timestamp information of the ground signal receiving equipment. The first mapping subunit is used to perform time-series dynamic compensation processing on the sampled sequence and its corresponding timestamp information. It analyzes the time-series pattern of the timestamp information through a recurrent neural network, estimates the symbol interval jitter caused by the Doppler effect, and uses an interpolation algorithm to resample the sampled sequence to obtain a time-aligned standardized sampled sequence. The second mapping subunit is used to perform multi-scale spectral feature extraction processing on the time-aligned standardized sampling sequence. It uses a parallel convolutional neural network with multi-sized convolutional kernels sliding in the time domain to extract subtle spectral fluctuations within a symbol period and capture the spectral evolution patterns over multiple symbol periods. It also fuses timestamp information to generate a spatiotemporally correlated feature map, thus obtaining a preliminary time-series spectral feature tensor. The third mapping subunit is used to perform parameterized feature enhancement processing on the preliminary time-series spectral feature tensor. It embeds the frequency reference parameter and quantization precision parameter as bias terms into the feature channel through a fully connected neural network, and performs weighted fusion of multi-scale features to obtain a unified high-dimensional time-frequency feature tensor. The first correction subunit is used to perform coarse equalization correction on the high-dimensional time-frequency feature tensor. It performs convolution operations in the feature space by sliding multi-size convolution kernels in a parallel convolutional neural network to extract local fine features and capture global features, and to compensate for amplitude distortion, thus obtaining a preliminary corrected feature tensor. The second correction subunit is used to refine the preliminary correction feature tensor through fine equalization. It learns the continuous phase error patterns in the preliminary correction feature tensor through a recurrent neural network and corrects the phase distortion point by point. It also performs dynamic weight adjustment in combination with the rapidly changing channel characteristics of the satellite link to obtain the refined equalization features. The third correction subunit is used to perform probabilistic symbol generation processing on the refined equilibrium features. It outputs the probability distribution of each symbol through a fully connected neural network and fuses the uncertainty estimate to obtain a preliminary equilibrium symbol representation with probabilistic features. The first estimation subunit is used to perform symbol timing offset estimation on the preliminary equalized symbol representation. It captures subtle timing jitter and identifies macroscopic temporal drift by sliding multi-size convolutional kernels on the symbol sequence through a parallel convolutional neural network, detects temporal irregularities at symbol boundaries, extracts timing error features by utilizing the adaptive learning capability of the convolutional neural network, and resamples the detected offset using an interpolation algorithm to obtain a preliminary timing-corrected symbol sequence. The second estimation subunit is used to perform joint estimation of phase and frequency deviation on the preliminary timing correction symbol sequence. It learns the Doppler spectrum hidden in the symbol sequence dynamically and compensates for phase rotation through a recurrent neural network to obtain the frequency deviation corrected symbol sequence. The third estimation subunit is used to perform feedback alignment optimization on the frequency offset corrected symbol sequence. It calculates the confidence weight of each position in the symbol sequence through the attention mechanism, identifies the symbol interval that needs to be realigned using the weighted features, and dynamically adjusts the symbol interval using an interpolation algorithm to obtain a time-aligned and phase-corrected symbol sequence. During the adjustment process, the adjustment result is backpropagated to the feature extraction layer through a closed-loop feedback mechanism. The first processing subunit is used to generate soft information vectors for the timing-aligned and phase-corrected symbol sequence. It converts the feature vector of each symbol in the symbol sequence into a probability distribution through a fully connected neural network and calculates the log-likelihood ratio to obtain the initial soft information vector. The second processing subunit is used to perform encoding constraint fusion processing on the initial soft information vector, and to iteratively update the soft information through a graph neural network to fuse the verification relationship of the satellite link LDPC code to obtain the enhanced soft information vector. The third processing subunit is used to integrate channel state information into the enhanced soft information vector. By dynamically weighting the channel statistical parameters, the weighted channel state information is fused with the enhanced soft information vector, and the output value is normalized to obtain the soft decision symbol set. The demodulation unit is used to iteratively refine the soft decision symbol set. It updates the uncertain region during feature reconstruction and performs multiple rounds of consistency verification by combining the LDPC code check relationship and channel state information to obtain the demodulated information.

Images