A channel decoding method and system based on dynamic channel environment

Abstract

This invention relates to the field of wireless communication channel decoding, specifically a channel decoding method and system based on dynamic channel environments. The method includes: acquiring a received signal; performing synchronization, channel estimation, and demodulation processing on the received signal to obtain an initial log-likelihood ratio vector and a channel-related feature matrix; concatenating and normalizing the initial log-likelihood ratio vector and the channel-related feature matrix to obtain a normalized feature matrix; constructing a log-likelihood ratio correction model; inputting the normalized feature matrix into the log-likelihood ratio correction model to obtain a corrected log-likelihood ratio vector; acquiring a pre-set initial value for a scaling factor; calibrating the corrected log-likelihood ratio vector to obtain a final log-likelihood ratio vector; and inputting the final log-likelihood ratio vector into a channel decoder for decoding. This method solves the problems of lack of adaptability and accuracy in existing channel decoding methods based on dynamic channel environments.

Description

Technical Field

[0001] This invention relates to the field of wireless communication channel decoding, specifically a channel decoding method and system based on dynamic channel environments. Background Technology

[0002] In scenarios such as near-Earth satellite communication, the channel state of wireless communication changes continuously over time. Traditional channel decoding methods are usually designed based on the quasi-static assumption of the channel, making it difficult to maintain stable decoding performance when the channel changes.

[0003] In existing technologies, channel estimation and decoding modules are independent. The soft information output from channel estimation is directly fed into the decoder. The decoding stage cannot utilize the temporal correlation information of the channel to correct this soft information. When the channel state changes, the decoder input mismatches with the actual channel state, leading to a decline in decoding performance. Furthermore, traditional decoding algorithms are based on ideal channel models, but real-world dynamic channels contain various non-ideal factors that are difficult to describe with precise mathematical models, resulting in limited performance of model-based designs in practical deployments. To address these issues, some research has attempted to introduce artificial intelligence to assist decoding; however, existing solutions are mostly designed for static scenarios, with limited adaptability to dynamic channel environments, and the models cannot be continuously updated after deployment to adapt to changes in channel characteristics.

[0004] Therefore, there is an urgent need for a channel decoding method that can adapt to dynamic channel environments and improve decoding accuracy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a channel decoding method and system based on dynamic channel environments, solving the problems of lack of adaptability and accuracy in existing channel decoding methods based on dynamic channel environments.

[0006] To achieve the above objectives, this invention provides a channel decoding method based on a dynamic channel environment. The method includes: acquiring a received signal; performing synchronization, channel estimation, and demodulation processing on the received signal to obtain an initial log-likelihood ratio vector and a channel-related feature matrix; concatenating and normalizing the initial log-likelihood ratio vector and the channel-related feature matrix to obtain a normalized feature matrix; constructing a log-likelihood ratio correction model; inputting the normalized feature matrix into the log-likelihood ratio correction model to obtain a corrected log-likelihood ratio vector; acquiring a pre-set initial value of a scaling factor; calibrating the corrected log-likelihood ratio vector to obtain a final log-likelihood ratio vector; and inputting the final log-likelihood ratio vector into a channel decoder for decoding.

[0007] This invention obtains an initial log-likelihood ratio vector and a channel-related feature matrix by synchronizing the received signal, estimating the channel, and demodulating the channel. This fully extracts the time-domain and state features of the channel, laying the foundation for subsequent corrections. By splicing and normalizing the vectors, a normalized feature matrix is ​​obtained, eliminating differences in feature dimensions and improving the effectiveness of the model input. By constructing a corrected model and inputting the normalized feature matrix, a corrected vector is obtained, achieving nonlinear optimization of soft information and adapting to dynamic channel characteristics. The final vector is obtained through scaling factor calibration, allowing the soft information to match the decoder input requirements. The decoder then completes the decoding, ensuring decoding accuracy and improving the adaptability to dynamic channels.

[0008] Optionally, the step of performing synchronization, channel estimation, and demodulation processing on the received signal to obtain an initial log-likelihood ratio vector and a channel-related feature matrix includes: performing carrier synchronization and symbol synchronization processing on the received signal to obtain a synchronized received signal; performing channel estimation on the synchronized received signal to obtain a channel estimation amplitude, a channel estimation phase, and a signal-to-noise ratio (SNR) estimate; demodulating the synchronized received signal and calculating an initial log-likelihood ratio vector by combining the channel estimation amplitude, the channel estimation phase, and the SNR estimate; extracting historical log-likelihood ratio vectors of adjacent code blocks of the synchronized received signal; and concatenating the historical log-likelihood ratio vectors, the channel estimation amplitude, the channel estimation phase, and the SNR estimate by channel into a multi-dimensional feature matrix to obtain a channel-related feature matrix.

[0009] This invention eliminates frequency offset, phase offset, and timing offset by synchronizing the received signal with carrier and symbol, ensuring the accuracy of subsequent signal processing. It extracts amplitude, phase, and signal-to-noise ratio through channel estimation to accurately capture the real-time channel status. It calculates the initial log-likelihood ratio vector by demodulation and combining it with the channel estimation value to obtain accurate initial soft information. It introduces channel timing correlation information by extracting the historical log-likelihood ratio vectors of adjacent code blocks. It obtains the channel-related feature matrix by splicing multiple feature channels, integrating multi-dimensional channel information and improving the accuracy of the channel-related feature matrix.

[0010] Optionally, the step of concatenating and normalizing the initial log-likelihood ratio vector and the channel-related feature matrix to obtain a normalized feature matrix includes: concatenating the initial log-likelihood ratio vector and the channel-related feature matrix by channel to form a multi-channel feature matrix; obtaining the pre-statistical global mean and global standard deviation; and using the global mean and global standard deviation to standardize the multi-channel feature matrix to obtain a normalized feature matrix.

[0011] This invention forms a multi-channel feature matrix by concatenating the initial log-likelihood ratio vector with the channel-related feature matrix, integrating soft information with multi-dimensional channel features to achieve comprehensive fusion of feature information. By retrieving the pre-statistical global mean and global standard deviation, and relying on the statistical characteristics of massive samples to ensure the scientific nature of standardization, the multi-channel feature matrix is ​​standardized using global statistics to obtain a normalized feature matrix. This eliminates the differences in the dimensions and numerical ranges of different features, giving each feature equal weight and improving the effectiveness and accuracy of the normalized feature matrix.

[0012] Optionally, the construction of the log-likelihood ratio correction model includes: obtaining the maximum Doppler frequency shift and symbol period of the target scene; calculating the coherence time based on the maximum Doppler frequency shift; calculating the number of symbols contained in the coherence time based on the symbol period; determining the total number of dilated convolutional layers based on the number of symbols; determining the dilation coefficient of each dilated convolutional layer based on the total number of layers and the number of symbols; constructing a dilated convolutional residual module based on the dilated convolutional layers and the dilation coefficients; and constructing the log-likelihood ratio correction model based on a pre-constructed normalization module and a dimension compression module combined with the dilated convolutional residual module.

[0013] This invention obtains the maximum Doppler frequency shift and symbol period of the target scene and calculates the coherence time and corresponding number of symbols, enabling the model to conform to the time-related characteristics of the dynamic channel. The number of symbols determines the number of dilated convolutional layers and the dilation coefficients, allowing the network's receptive field to accurately cover the channel's coherence time range. By building a dilated convolutional residual module, it effectively extracts multi-scale temporal features of the channel and alleviates gradient vanishing. By combining normalization and dimensionality compression modules to build a corrected model, the model training becomes more stable and outputs a single-channel vector adapted for decoding, improving the model's ability to extract channel time-related features and enhancing the performance of the log-likelihood ratio corrected model.

[0014] Optionally, obtaining a preset initial value for the scaling factor and calibrating the corrected log-likelihood ratio vector to obtain the final log-likelihood ratio vector includes: obtaining a preset initial value for the scaling factor; initially adjusting the corrected log-likelihood ratio vector to obtain an initially adjusted log-likelihood ratio vector; performing decoding processing and cyclic redundancy check based on the initially adjusted log-likelihood ratio vector, and then calculating the decoding accuracy using a sliding window; when the decoding accuracy is lower than a preset baseline accuracy, calculating the difference between the decoding accuracy and the baseline accuracy; adjusting the initial value for the scaling factor based on the difference to obtain an updated scaling factor; and calculating the final log-likelihood ratio vector based on the updated scaling factor.

[0015] This invention initially adjusts the corrected log-likelihood ratio vector by setting a preset scaling factor initial value, allowing soft information to match the decoder input characteristics first. Through decoding and cyclic redundancy checks, the validity of the decoding result is accurately judged. Combined with sliding window statistics of decoding accuracy, decoding performance is monitored in real time. By comparing the difference with the benchmark accuracy, the degree of performance deviation is quantified. The scaling factor is adjusted based on the difference to obtain the final log-likelihood ratio vector, realizing dynamic optimization of the scaling factor and improving the scientific nature of the final log-likelihood ratio vector.

[0016] Optionally, obtaining the preset initial value of the scaling factor includes: obtaining the channel parameters of the target dynamic channel scenario; generating a simulated received signal and corresponding real transmitted bits based on the channel parameters; processing the simulated received signal using the log-likelihood ratio correction model to obtain a simulated corrected log-likelihood ratio vector; setting a candidate value range for the scaling factor; traversing candidate scaling factors within the candidate value range using a grid search algorithm; calibrating and decoding the simulated corrected log-likelihood ratio vector using the candidate scaling factors to obtain simulated decoding; obtaining the simulated decoding accuracy corresponding to the candidate scaling factor based on the simulated decoding and the real transmitted bits; calculating the average decoding accuracy of each candidate scaling factor; and selecting the candidate scaling factor with the highest average decoding accuracy as the initial value of the scaling factor.

[0017] This invention generates simulated signals and real bits by acquiring target channel parameters, allowing scaling factor calibration to fit the actual channel scenario. It obtains a simulated correction vector by processing the simulated signal through a correction model, ensuring consistency between the calibration object and actual inference. It achieves accurate optimization across the entire range by traversing candidate scaling factor values ​​through grid search. It calibrates decoding through candidate factors and calculates the accuracy rate, quantifying the adaptation effect of each factor. By selecting the optimal factor as the initial value, the initial calibration of soft information matches the optimal input of the decoder. This invention achieves scenario-based and accurate selection of the initial value of the scaling factor, avoiding decoding performance loss caused by blind setting, and improving the adaptation between soft information and the decoder after initial calibration.

[0018] Optionally, adjusting the initial value of the scaling factor based on the difference to obtain the updated scaling factor includes: obtaining a data pair of deviation and adjustment values ​​through simulation; fitting the data pair of deviation and adjustment values ​​to obtain a deviation adjustment relationship model; and inputting the difference value into the deviation adjustment relationship model to obtain the updated scaling factor.

[0019] This invention obtains data pairs of deviation and adjustment values ​​through simulation, providing a quantitative basis for scaling factor adjustment that closely matches the actual channel. By fitting the data pairs, a deviation adjustment relationship model is obtained, establishing a precise mapping relationship between performance deviation and factor adjustment. By inputting the difference into the model, the updated scaling factor is obtained, achieving rapid and quantitative adjustment of the scaling factor. Overall, this avoids the blindness and trial-and-error of factor adjustment, improving the efficiency and accuracy of dynamic scaling factor adjustment.

[0020] Optionally, the method further includes: obtaining information bits based on the decoding process, and performing cyclic redundancy check on the information bits to obtain a high-confidence decoding result; associating the high-confidence decoding result with the corresponding initial log-likelihood ratio vector and the channel correlation feature matrix to form incremental training samples; and updating the log-likelihood ratio correction model using the incremental training samples.

[0021] This invention performs cyclic redundancy check on the decoded information bits to select high-confidence decoding results, ensuring the reliability of the samples. By associating high-confidence results with the corresponding initial log-likelihood ratio vector and channel-related feature matrix to form incremental training samples, the samples are made to fit the real-time characteristics of the actual channel. The log-likelihood ratio correction model is updated by incremental training samples, enabling the model to continuously adapt to the dynamic channel characteristics and improve the model's adaptability and long-term stability to dynamic channels.

[0022] Optionally, the step of associating the high-confidence decoding result with the corresponding initial log-likelihood ratio vector and the channel correlation feature matrix to form incremental training samples includes: obtaining a pseudo-label log-likelihood ratio based on the high-confidence decoding result and the received signal; smoothing the pseudo-label log-likelihood ratio to obtain a smoothed pseudo-label; and associating the smoothed pseudo-label with the corresponding initial log-likelihood ratio vector and the channel correlation feature matrix to form labeled training samples.

[0023] This invention obtains the pseudo-label log-likelihood ratio through high-confidence decoding results and received signals, providing a supervision target that fits the actual channel for incremental training. By smoothing the pseudo-labels, the label error caused by noise is reduced, and the label reliability is improved. By associating the smoothed pseudo-labels with the corresponding initial log-likelihood ratio vector and the channel-related feature matrix to form labeled training samples, the accurate matching between input features and supervision targets is achieved, thereby improving the effectiveness and robustness of incremental training samples.

[0024] Another aspect of the present invention provides a channel decoding system based on a dynamic channel environment, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the channel decoding method based on a dynamic channel environment as described in any of the preceding aspects of the present invention.

[0025] The present invention provides a channel decoding system based on a dynamic channel environment, which is compact, stable, highly integrated, and simple in construction. It can stably execute the channel decoding method based on a dynamic channel environment provided in the preceding part of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of a channel decoding method based on a dynamic channel environment according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a channel decoding system structure based on a dynamic channel environment according to an embodiment of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0028] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0029] Please see Figure 1 To address the technical problems, in one alternative embodiment, such as Figure 1 The channel decoding method based on a dynamic channel environment, as shown, includes the following steps: Step S1: Acquire the received signal, perform synchronization, channel estimation and demodulation processing on the received signal to obtain the initial log-likelihood ratio vector and the channel correlation feature matrix.

[0030] In this embodiment, radio frequency signals from near-Earth satellites are captured by a receiving antenna. These signals are amplified by a low-noise amplifier, converted to intermediate frequency (IF) signals by a down-converter, and then sampled by an analog-to-digital converter to obtain digital IF signals. The digital IF signals are then digitally down-converted to a baseband complex signal sequence. This baseband complex signal sequence is the received signal, with each sampling point corresponding to a complex value representing the in-phase and quadrature components, respectively. The sampling rate of the received signal is matched to the symbol period; in this embodiment, the symbol period is 1 μs, corresponding to a sampling rate of 1 MHz.

[0031] The process of synchronizing, estimating, and demodulating the received signal to obtain the initial log-likelihood ratio vector and the channel correlation feature matrix specifically includes the following sub-steps: Step S101: Perform carrier synchronization and symbol synchronization processing on the received signal to obtain a synchronized received signal.

[0032] In this embodiment, the received signal is a baseband complex signal sequence. First, carrier synchronization processing is performed on the received signal: frequency offset estimation is performed using a preamble sequence or pilot symbols, and the frequency difference between the received signal and the local reference signal is calculated using a Fast Fourier Transform (FFT) to obtain a Doppler frequency offset estimate. Then, a complex exponential signal with the opposite phase to the frequency offset estimate is generated by a numerically controlled oscillator (NCO). The received signal is multiplied by this complex exponential signal to complete frequency offset compensation. Subsequently, a phase-locked loop (PLL) is used to track the residual phase offset to achieve phase offset compensation, resulting in a carrier-synchronized signal. Next, symbol synchronization processing is performed on the carrier-synchronized signal: the Gardner timing error detection algorithm is used to calculate the timing error within each symbol period. The sampling time is adjusted using an interpolation filter (such as cubic interpolation) to align the sampling point with the optimal decision time of the eye diagram, outputting a synchronized received signal with a complex value for each symbol period. This signal eliminates the effects of Doppler frequency shift, carrier phase offset, and timing offset, providing an accurate baseband signal sequence for subsequent channel estimation.

[0033] Step S102: Perform channel estimation on the synchronized received signal to obtain the channel estimation amplitude, channel estimation phase and signal-to-noise ratio estimation value.

[0034] In this embodiment, channel estimation is performed using pilot symbols embedded in the synchronized received signal. First, pilot symbols are extracted from the synchronized received signal, denoted as... ,in , This represents the total number of pilot symbols. These pilot symbols are reference symbols known to the receiver, denoted as... The channel estimate for the pilot position is calculated using the least squares algorithm: For data symbols at non-pilot positions, a linear interpolation method is used to calculate the channel estimate for each symbol position based on the channel estimates for adjacent pilot positions, resulting in a complex channel coefficient vector. Its length is the code block length. , No. The channel estimate for each symbol is denoted as Then, the magnitude of the complex channel coefficient vector is extracted. and phase These constitute the channel estimation magnitude vector and the channel estimation phase vector, respectively, both with a length of [missing information]. .

[0035] Simultaneously, the noise variance is calculated using the reception error of the pilot symbols: the error is calculated at the pilot position. The variance of the pilot position error is statistically analyzed, and then the noise variance estimate for each symbol position is obtained by interpolation, forming a noise variance vector. , its first The elements are denoted as , indicating the first Noise power at each symbol position.

[0036] The signal-to-noise ratio estimate is obtained by converting the noise variance vector: The first... The signal-to-noise ratio estimate for each symbol position is: This forms a signal-to-noise ratio estimate vector with a length of [value missing]. This signal-to-noise ratio estimate vector will be used as input for subsequent feature concatenation to characterize the channel quality at each symbol location.

[0037] Step S103: Demodulate the synchronized received signal and calculate the initial log-likelihood ratio vector by combining the channel estimation amplitude, the channel estimation phase, and the signal-to-noise ratio estimate.

[0038] In this embodiment, QPSK modulation is used, with each symbol carrying 2 bits. For the first received signal after synchronization... A symbol, denoted as The corresponding channel estimate is ,in The channel estimation amplitude obtained in step S102, To estimate the phase of the channel, The imaginary unit satisfies The noise variance at the symbol location is The noise variance vector output from step S102 is provided. For QPSK modulation, each symbol corresponds to two bits, denoted as follows: and . No. The initial log-likelihood ratio of the two bits of a symbol is calculated using the following formula: First, channel equalization is performed on the received symbols to obtain the equalized symbols. ,in For in-phase components, These are orthogonal components. Then the initial log-likelihood ratio of the first bit is... for The initial log-likelihood ratio of the second bit for Iterate through all cases using this formula. For each symbol, calculate the log-likelihood ratio of its two bits, and arrange them in symbol and bit order to form a sequence of length [length missing]. The initial log-likelihood ratio vector Each element in this vector represents the soft information of the corresponding bit. A positive value indicates that the bit tends to be 1, a negative value indicates that it tends to be 0, and the absolute value indicates the level of confidence.

[0039] Step S104: Extract the historical log-likelihood ratio vector of adjacent code blocks of the synchronized received signal.

[0040] In this embodiment, the receiver maintains a historical log-likelihood ratio (LPR) buffer queue to store the corrected LPR vector of the previous code block. For the currently processed code block, the corrected LPR vector of the previous code block is directly read from this buffer queue as the historical LPR vector. If the current code block is the first frame of transmission, the buffer queue is empty. In this case, the historical LPR vector is set to zero, with a length of 4096 (two bits per symbol), the same as the initial LPR vector. This historical vector reflects the channel soft information characteristics of the most recent code block, providing timing correlation information for subsequent feature concatenation. After the corrected LPR vector of the current code block is calculated, it is immediately stored in the buffer queue as the historical LPR vector for the next code block.

[0041] Step S105: The historical log-likelihood ratio vector, the channel estimation amplitude, the channel estimation phase, and the signal-to-noise ratio estimation value are concatenated by channel into a multi-dimensional feature matrix to obtain the channel-related feature matrix.

[0042] In this embodiment, the historical log-likelihood ratio vector (length 4096), channel estimation amplitude vector (length 2048), channel estimation phase vector (length 2048), and signal-to-noise ratio (SNR) estimation vector (length 2048) are first unified to the same length. Since each symbol corresponds to two bits, the channel estimation amplitude vector, channel estimation phase vector, and SNR estimation vector are each copied and extended to a length of 4096, so that each bit position corresponds to its respective channel amplitude, channel phase, and SNR estimate. Then, these four vectors, each of length 4096, are concatenated side-by-side along the channel dimension to form a 4096-row, 4-column multidimensional feature matrix. This matrix is ​​the channel-related feature matrix, where each column corresponds to a feature type, and each row corresponds to all channel-related features for one bit position.

[0043] Step S2: The initial log-likelihood ratio vector and the channel-related feature matrix are concatenated and normalized to obtain a normalized feature matrix.

[0044] Specifically, the initial log-likelihood ratio vector and the channel-related feature matrix are concatenated and normalized to obtain a normalized feature matrix, which includes: Step S201: The initial log-likelihood ratio vector and the channel-related feature matrix are concatenated by channel to form a multi-channel feature matrix.

[0045] In this embodiment, the initial log-likelihood ratio vector has a length of 4096, corresponding to the soft information value of each bit, and is considered as a single-channel feature vector. The channel-related feature matrix output in step S105 has a dimension of 4096 rows by 4 columns, where each column corresponds to four features: historical log-likelihood ratio, channel estimated amplitude, channel estimated phase, and signal-to-noise ratio estimate, and each row corresponds to a bit position. The initial log-likelihood ratio vector is used as the fifth channel and concatenated with the four channels of the channel-related feature matrix in terms of channel dimension, that is, keeping the number of rows (bit positions) unchanged and adding a feature channel. In specific operation, the initial log-likelihood ratio vector is added as a new column to the right side of the channel-related feature matrix to form a new multi-dimensional feature matrix. The dimension of this matrix is ​​4096 rows by 5 columns, where the first to fourth columns are the historical log-likelihood ratio, channel estimated amplitude, channel estimated phase, and signal-to-noise ratio estimate, respectively, and the fifth column is the initial log-likelihood ratio of the current bit. This matrix is ​​the multi-channel feature matrix. Each row corresponds to a multi-dimensional feature vector of a bit position, which contains the current soft information, historical soft information and channel state information of the current symbol for that bit, providing a structured feature representation for subsequent normalization processing and neural network input.

[0046] Step S202: Obtain the pre-statistical global mean and global standard deviation.

[0047] In this embodiment, a simulation training dataset covering typical channel conditions for near-Earth satellite communication is first constructed. This dataset contains multiple code block samples, each processed through steps S101 to S201 to obtain a corresponding multi-channel feature matrix. The dataset generation fully considers the channel variation range: signal-to-noise ratio covers -5dB to 20dB, and Doppler shift covers 0 to 50kHz, ensuring that the training data reflects the channel statistical characteristics of the target scenario. Then, a channel-by-channel statistical analysis is performed on the multi-channel feature matrices of all samples in the training set. Specifically, for each of the five feature channels, the arithmetic mean of all values ​​at all positions of all samples in that channel is calculated as the global mean of that channel; simultaneously, the sample standard deviation of all values ​​of all samples in that channel is calculated as the global standard deviation of that channel. Since the physical meaning and numerical range of each channel are different, the five channels are statistically analyzed independently, resulting in five global means and five global standard deviations. After calculation, these five global means and five global standard deviations are stored as constants in the non-volatile memory of the receiving device for direct retrieval during the online inference phase. Once determined, this statistic remains unchanged during subsequent online deployments to ensure consistency in feature preprocessing between the training and inference phases. When the channel parameters of the target communication scenario (such as Doppler shift range and signal-to-noise ratio range) undergo significant changes, a new global statistic adapted to the new scenario must be generated and updated to non-volatile memory.

[0048] Step S203: Standardize the multi-channel feature matrix using the global mean and global standard deviation to obtain a normalized feature matrix.

[0049] In this embodiment, five pre-statistical and stored global means and five global standard deviations are read from non-volatile memory, corresponding to the five feature channels: historical log-likelihood ratio, channel estimated amplitude, channel estimated phase, signal-to-noise ratio estimate, and initial log-likelihood ratio, respectively. Simultaneously, the multi-channel feature matrix formed in step S201 is obtained. This matrix has 4096 rows, each corresponding to one bit position, and 5 columns, each corresponding to one feature channel. Standardization processing is performed independently for each channel: for the first feature channel (historical log-likelihood ratio), all 4096 values ​​of this channel are extracted, and each value is subtracted from the corresponding global mean and then divided by the corresponding global standard deviation; for the second feature channel (channel estimated amplitude), the same operation is performed, using the corresponding global mean and global standard deviation; and so on, processing the third feature channel (channel estimated phase), the fourth feature channel (signal-to-noise ratio estimate), and the fifth feature channel (initial log-likelihood ratio) in sequence. After all five channels have been processed, a new matrix of 4096 rows by 5 columns is obtained, which is the normalized feature matrix. Through the above standardization process, the numerical distribution of each feature channel is adjusted to have a mean of 0 and a standard deviation of 1, eliminating differences in dimensions and numerical ranges between different features, thus ensuring that each feature channel has equal importance during subsequent neural network training and inference.

[0050] Step S3: Construct a log-likelihood ratio correction model by inputting the normalized feature matrix into the log-likelihood ratio correction model to obtain the corrected log-likelihood ratio vector.

[0051] The construction of the log-likelihood ratio correction model specifically includes the following sub-steps: Step S301: Obtain the maximum Doppler frequency shift and symbol period of the target scene.

[0052] In this embodiment, the target communication scenario is a near-Earth satellite communication system, whose system parameters have been determined during the satellite communication link design phase. The orbital altitude is 500 km, the carrier frequency is 2 GHz, and based on the relative speed between the satellite and the ground terminal, the maximum Doppler shift is calculated to be 50 kHz. This maximum Doppler shift is an extreme value occurring when the satellite is at a low elevation angle and its motion direction is consistent with the propagation direction, serving as the worst-case operating condition parameter for the system design. The symbol period is determined by the system bandwidth; this embodiment uses a 1 MHz signal bandwidth, corresponding to a symbol period of 1 microsecond. This value is jointly determined by the analog-to-digital converter sampling rate and the pulse shaping filter. Both of these parameters are fixed configuration parameters during system design, read from the configuration file during receiver initialization, or stored in the receiver's non-volatile memory. After obtaining these two parameters, they will be used for subsequent channel coherence time calculation and temporal convolutional network structure design.

[0053] Step S302: Calculate the coherence time based on the maximum Doppler frequency shift.

[0054] In this embodiment, the coherence time is calculated based on the maximum Doppler frequency shift. The coherence time is inversely proportional to the maximum Doppler frequency shift; that is, the larger the maximum Doppler frequency shift, the faster the channel changes, and the shorter the coherence time. In this embodiment, the maximum Doppler frequency shift is 50 kHz, corresponding to a coherence time of approximately 20 microseconds. This coherence time represents the length of time during which the channel impulse response remains essentially unchanged, providing a basis for subsequently determining the receptive field of the temporal convolutional network.

[0055] Step S303: Calculate the number of symbols contained in the coherence time based on the symbol period.

[0056] In this embodiment, calculating the number of symbols included in the coherence time based on the symbol period can be done by dividing the coherence time by the symbol period, i.e. ,in, , ,get This indicates that the channel remains essentially unchanged over 20 symbol periods. This value reflects the correlation length of the channel in the time dimension, that is, the range of channel changes that the sequential convolutional network needs to capture. This number of symbols will serve as an important basis for subsequently determining the number of dilated convolutional layers and dilation coefficients, ensuring that the network's receptive field can cover the time window in which the channel remains relevant, thereby effectively extracting the temporal correlation features of the channel.

[0057] Step S304: Determine the total number of dilated convolutional layers based on the number of symbols.

[0058] In this embodiment, in order for the temporal convolutional network to effectively capture the changing characteristics of the channel during the coherence time, the total receptive field of the dilated convolutional residual module must be no less than the number of symbols contained in the coherence time. The formula for calculating the receptive field of dilated convolution is: in The kernel size is [size]. To increase the number of convolutional layers, For the first The expansion coefficients of the layers are arranged in a geometrically increasing sequence.

[0059] Quantitative calculations show that: If 3 layers are taken, the expansion coefficient sequence is as follows: Always feel wild If the value is less than 20, it cannot fully cover the symbol range of the channel coherence time, resulting in the loss of some timing features; If 4 layers are selected, the expansion coefficient sequence is as follows: Always feel wild With a value greater than 20, it can fully cover the symbol range of coherent time and retain a redundancy of 11 symbols, adapting to the burst time-varying characteristics of near-Earth satellite channels; If 5 layers are selected, the expansion coefficient sequence is as follows: Always feel wild Although the coverage is more comprehensive, the number of model parameters has increased from 623,000 in 4 layers to 1,247,000 (doubled), and the single-frame inference time has increased from 8μs to 18μs, exceeding the real-time processing constraints of the near-Earth satellite receiver (edge ​​device) (requiring a single-frame inference time of ≤10μs).

[0060] Therefore, considering the receptive field coverage requirements, the redundancy requirements of dynamic channel bursts and time variations, and the computing power and real-time constraints of edge devices, choosing a 4-layer dilated convolutional structure is the optimal solution that balances performance, scenario adaptability, and engineering feasibility.

[0061] Step S305: Determine the dilation coefficient of each of the dilated convolutional layers based on the total number of layers and the number of symbols.

[0062] In this embodiment, the receptive field size of the dilated convolutional layer is determined by both the kernel size and the dilation coefficients of each layer. To ensure the network can effectively capture channel changes during the coherence time, the total receptive field must be at least 20, the number of symbols contained within the coherence time. With a fixed kernel size of 3 and a fixed number of layers of 4, and after dynamic channel feature adaptation verification, a doubling sequence of dilation coefficients is selected. Instead of an arithmetic or random sequence, the reason is as follows: Quantization of receptive field adaptation: The total receptive field of this sequence is 31, which can cover 20 coherent symbols. The receptive field of each layer is gradually expanded to 3, 7, 15 and 31, realizing progressive feature extraction from short time to long time, which is well adapted to the time-varying characteristics of near-Earth satellite channels, which are "slow changes are the main feature and sudden fast changes are the auxiliary feature". Multi-scale temporal feature extraction: Small expansion coefficients (1, 2) capture short-term channel correlation (1-7 symbols), adapting to the sudden rapid changes in the channel; large expansion coefficients (4, 8) capture long-term channel dependence (15-31 symbols), adapting to the overall slow changes in the channel. Multi-scale feature fusion can improve the model's adaptability to dynamic channels. Engineering feasibility: The expansion coefficient of the doubling sequence is a power of 2, which can be implemented in hardware (FPGA / MCU) through shift operations without multiplication, reducing the computational complexity of edge devices and improving real-time performance.

[0063] The total receptive field corresponding to this sequence is 31, which is greater than the number of symbols (20) contained in the coherence time. This sufficiently covers the time window of channel maintenance correlation, ensuring that the network effectively learns the temporal evolution of the channel. At the same time, the computational cost is moderate, making it suitable for deployment on edge devices. Once the expansion coefficient is determined, it is fixed during the model training phase and will not be adjusted during the inference phase.

[0064] Step S306: Construct an dilated convolutional residual module based on the dilated convolutional layer and the dilation coefficient.

[0065] In this embodiment, the dilated convolutional residual module consists of four stacked residual blocks. Each residual block contains a dilated convolutional layer, an activation function layer, and a residual connection. All dilated convolutional layers use the same padding (SamePadding) to ensure that the sequence length remains constant at 4096 lines.

[0066] The first residual block uses an expanded convolutional layer with an expansion factor of 1, a kernel size of 3, 5 input channels, and 64 output channels. After extracting temporal features from the input feature matrix (extracting time-related features in the bit / symbol dimension to adapt to the time evolution of dynamic channels), it is activated by the ReLU activation function. Then, the activated output is residually concatenated with the input (since the number of input channels is 5 and the number of output channels is 64, the number of channels is mismatched, so a 1×1 convolution is used to adjust the number of input channels to 64 before adding them), thus obtaining the output of the first residual block.

[0067] The second residual block takes the output of the first residual block as input, uses an expanded convolutional layer with an expansion factor of 2, a kernel size of 3, 64 input channels, and 64 output channels. After ReLU activation, it is residually connected to the input (at this time, both the input and output channels have 64 channels, so they are directly added together) to obtain the output of the second residual block.

[0068] The third residual block takes the output of the second residual block as input, uses an expanded convolutional layer with an expansion factor of 4, a kernel size of 3, 64 input channels, and 64 output channels, and after ReLU activation, it is residually connected to the input to obtain the output of the third residual block.

[0069] The fourth residual block takes the output of the third residual block as input, uses an expanded convolutional layer with an expansion factor of 8, a kernel size of 3, 64 input channels, and 64 output channels. After ReLU activation, it is residually connected to the input to obtain the output of the fourth residual block.

[0070] The input to this module is the normalized feature matrix output from step S203, with dimensions of 4096 rows by 5 columns. After processing through 4 residual blocks, the output is a feature map with dimensions of 4096 rows by 64 columns. Residual connections can alleviate the gradient vanishing problem in deep networks, ensuring that the model can effectively learn the long-term temporal evolution of dynamic channels. This module is the core component for the model to extract the temporal features of dynamic channels.

[0071] Step S307: Construct a log-likelihood ratio correction model based on the pre-built normalization module and dimension compression module combined with the dilated convolution residual module.

[0072] In this embodiment, the normalized feature matrix is ​​used as the model input. Based on the temporal convolutional network (TCN) architecture, a complete log-likelihood ratio correction model is constructed by sequentially passing through a normalization module, a dilated convolutional residual module, and a dimension compression module.

[0073] It should be noted that the standardization process in step S203 is a global scale unification of the entire dataset, while layer normalization is a channel dimension normalization within each sample. The two have different functions, but both adapt to the time-varying characteristics of dynamic channels: The near-Earth satellite channels targeted by this invention are dynamic time-varying channels, and the distribution of sample features changes in real time with the channel state. Batch normalization (BatchNorm) relies on the statistical characteristics of batch samples, which will cause normalization to fail under dynamic channels due to the time-varying distribution. Layer normalization (LayerNorm) normalizes each sample independently, without relying on batch statistics, and can adapt to the time-varying characteristics of sample distribution in dynamic channels, making model training more stable.

[0074] First, the normalization module uses layer normalization to normalize the input feature matrix. This process is performed independently along the feature channel dimension, making the feature distribution of each sample tend to be stable. Compared with batch normalization, it is more suitable for sequence tasks and is not sensitive to the size of small batches. The output dimension remains unchanged, still 4096 rows by 5 columns.

[0075] Subsequently, the normalized feature matrix is ​​input into the dilated convolutional residual module constructed in step S306. This module consists of four stacked residual blocks with dilation coefficients of 1, 2, 4, and 8, respectively. The kernel size is 3 for each residual block, and the number of channels is gradually increased from 5 to 64. The dilated convolution extracts temporal features at different time scales, and the residual connections alleviate the gradient vanishing problem in deep networks. The output is a feature map with dimensions of 4096 rows by 64 columns.

[0076] Finally, the aforementioned feature map is input into a dimension compression module, which employs a 1×1 convolutional layer with a kernel size of 1, 64 input channels, and 1 output channel. The core principle of this design is to adapt to the decoder's input requirements: the channel decoder only receives single-channel log-likelihood ratio (LLR) vectors, rather than multi-channel feature maps. By linearly combining the channel dimensions at each position of the feature map using a 1×1 convolution, the 64-channel temporal features are compressed into a single-channel LLR vector. This preserves the temporal feature correction results of the dynamic channel while perfectly matching the decoder's input format.

[0077] The output is a corrected log-likelihood ratio vector with 4096 rows and 1 column. At this point, the complete log-likelihood ratio correction model is built. This model can map the input normalized feature matrix to a corrected log-likelihood ratio vector, achieving nonlinear enhancement of the initial soft information. Furthermore, all module designs are adapted to the time-varying characteristics of near-Earth satellite dynamic channels and the engineering implementation constraints of edge devices.

[0078] The log-likelihood ratio correction model satisfies the following formula: in, For the total loss function, As the baseline loss function, This is the dynamic channel weighted penalty coefficient. The sign consistency penalty coefficient, This is a dynamic channel weighted penalty term. This is a symbol consistency penalty term. and The determined penalty coefficient can be optimized offline on the validation set using grid search.

[0079] in, This represents the total number of bits in a single code block. For the first The corrected log-likelihood ratio for each bit. For the first The log-likelihood ratio of the pseudo-labels for each bit.

[0080] The baseline loss function is used to constrain the corrected log-likelihood ratio and the numerical fitting accuracy of the pseudo-label.

[0081] in, To prevent small constants from being divided by zero, For the first The signal-to-noise ratio estimate for each bit. For the first Channel estimation amplitude of 1 bit.

[0082] The dynamic channel weighted penalty term addresses the core challenge of non-uniform channel conditions in dynamic channel scenarios. Traditional benchmark MSE loss imposes an equal penalty on the error at all bit positions, failing to differentiate between favorable and unfavorable channel conditions. This results in insufficient optimization priority for adverse channel positions such as low SNR and weak channel amplitude, becoming a bottleneck for decoding performance in dynamic channels. Therefore, an adaptive channel state weighting factor is introduced. When the channel conditions are worse ( The lower, The smaller the weight factor, the larger the error penalty at the corresponding bit position is automatically amplified, forcing the model to prioritize optimizing the LLR correction accuracy under harsh dynamic channels during training, fundamentally improving the robustness of the model under time-varying channels, and completely solving the industry pain point of the traditional model's performance plummeting when the channel deteriorates.

[0083] The sign consistency penalty term addresses the core characteristic of LDPC decoding: its extreme sensitivity to the log-likelihood ratio sign. The core logic of LDPC decoding determines the 0 / 1 value of a bit based on the sign of the LLR. Sign errors directly lead to decoding failure. Traditional benchmark MSE loss only constrains the numerical fitting accuracy of the LLR, not directly constraining sign consistency, easily resulting in situations where numerical fitting is good but sign errors occur, severely impacting decoding performance. Furthermore, traditional sign penalties based on indicator functions are undifferentiable and cannot be used for deep learning backpropagation. This solution constructs a continuously differentiable sign consistency constraint based on cosine similarity: when the predicted LLR and the pseudo-label LLR have the same sign, their product is positive, the cosine similarity is 1, and the penalty term is 0; when the signs are opposite, the product is negative, the cosine similarity is -1, and the penalty term reaches its maximum value of 2, achieving a smooth penalty for sign errors. This form directly addresses the core requirement of LDPC decoding by constraining sign correctness and perfectly adapts to the backpropagation training process of deep learning, fundamentally reducing the sign error rate and significantly improving decoding accuracy.

[0084] Within a batch For each sample, calculate... Then, the average value is taken as the final loss for that batch: Finally, during prediction, the normalized feature matrix is ​​input into the trained log-likelihood ratio correction model to obtain the corrected log-likelihood ratio vector.

[0085] Step S4: Obtain the initial value of the preset scaling factor, and perform calibration processing on the corrected log-likelihood ratio vector to obtain the final log-likelihood ratio vector.

[0086] The process of obtaining a pre-set initial value for the scaling factor and calibrating the corrected log-likelihood ratio vector to obtain the final log-likelihood ratio vector includes: Step S401: Obtain the preset initial value of the scaling factor, and perform initial adjustment on the corrected log-likelihood ratio vector to obtain the initially adjusted log-likelihood ratio vector.

[0087] The specific steps for obtaining the initial value of the preset scaling factor include the following: Step S40101: Obtain the channel parameters of the target dynamic channel scenario, and generate the simulated received signal and the corresponding real transmitted bits based on the channel parameters.

[0088] In this embodiment, the system configuration parameters for the near-Earth satellite communication scenario, namely the channel parameters, are first obtained, including a maximum Doppler shift of 50 kHz, a symbol period of 1 microsecond, a carrier frequency of 2 GHz, a sampling rate of 1 MHz, a modulation scheme of QPSK, a channel coding scheme of LDPC code with a code rate of 1 / 2, and a code block length of 2048 symbols. A channel simulation model is constructed based on these channel parameters: the transmitter randomly generates an information bit sequence, which is then LDPC encoded and QPSK modulated to obtain a baseband transmitted symbol sequence. The channel model uses a frequency-flat Rayleigh fading channel. A time-varying channel coefficient sequence is generated based on the maximum Doppler shift, with the rate of change of the channel coefficients matching the 50 kHz Doppler shift. The transmitted symbols are multiplied by the corresponding channel coefficients, and additive white Gaussian noise is added. The noise power is determined according to a preset signal-to-noise ratio (SNR), with the SNR coverage range set to -5 dB to 20 dB and a step size of 1 dB. 500 frames of data are generated independently for each SNR point. The receiver receives the simulated signal and records the actual transmitted bits corresponding to each frame. The above process generates a simulation dataset containing multiple signal-to-noise ratio points. Each dataset contains a simulated received signal sequence and a corresponding real transmitted bit sequence, which are used for offline calibration of the initial value of the scaling factor.

[0089] Step S40102: Process the simulated received signal using the log-likelihood ratio correction model to obtain the simulated corrected log-likelihood ratio vector.

[0090] In this embodiment, the simulated received signal is first synchronized with both carrier and symbol to obtain a synchronized simulated signal. Then, channel estimation is performed on the synchronized simulated signal to extract the estimated channel amplitude, estimated channel phase, and noise variance vector. Next, demodulation is performed and the initial log-likelihood ratio vector is calculated. Simultaneously, the corrected log-likelihood ratio vector from the previous frame is extracted as historical features. The initial log-likelihood ratio vector is concatenated with the historical log-likelihood ratio vector, the estimated channel amplitude, the estimated channel phase, and the noise variance vector, and then normalized to obtain a normalized feature matrix. Finally, this normalized feature matrix is ​​input into the log-likelihood ratio correction model for forward propagation calculation. The model output is the simulated corrected log-likelihood ratio vector, which has the same physical meaning and numerical range as the corrected log-likelihood ratio vector obtained during online inference and is used for offline calibration of the initial value of the scaling factor.

[0091] Step S40103: Set the candidate value range of the scaling factor, and use a grid search algorithm to traverse the candidate scaling factors within the candidate value range.

[0092] In this embodiment, magnitude analysis is performed on all simulated and corrected log-likelihood ratio vectors to calculate the distribution range of their absolute mean. The initial lower bound of the candidate scaling factor range is set to 0.2, and the initial upper bound is set to 5.0 to ensure coverage of possible scaling requirements. A two-stage grid search strategy is adopted: The first stage is a coarse search, which generates a set of 25 candidate scaling factors in the range of 0.2 to 5.0 with a step size of 0.2. The second stage is a fine search, which determines the suboptimal interval that minimizes the bit error rate based on the coarse search results. Within this interval, a fine-grained set of candidate scaling factors is generated with a step size of 0.05. If the coarse search finds that the optimal value falls on the search boundary, the boundary is expanded outward by 0.2 (for example, the lower bound is lowered to 0.1 or the upper bound is raised to 5.2), and the coarse search is repeated until the optimal value does not fall on the boundary. During the grid search process, each candidate scaling factor is sequentially retrieved, and the simulation-corrected log-likelihood ratio vector is calibrated by dividing the corrected log-likelihood ratio vector by the candidate scaling factor. This division operation can match the soft information amplitude with the decoder input expectation: when the decoder input LLR amplitude is too large, the amplitude is reduced by increasing the scaling factor; when it is too small, the scaling factor is increased by decreasing the scaling factor, thereby optimizing the decoding performance.

[0093] Step S40104: The simulated modified log-likelihood ratio vector is calibrated and decoded using the candidate scaling factors to obtain the simulation decoding.

[0094] In this embodiment, firstly, for each frame of the simulated corrected log-likelihood ratio vector, each element is divided by the current candidate scaling factor to obtain the calibrated simulated log-likelihood ratio vector. This division operation is used to adjust the amplitude of the soft information so that the distribution of the soft information matches the input characteristics desired by the decoder. Then, the calibrated simulated log-likelihood ratio vector is input into the same LDPC decoder deployed online. The decoder uses the belief propagation algorithm, with a fixed number of iterations of 5, and outputs the decoded information bit sequence as the simulated decoding result for that frame. The above process is performed independently for each candidate scaling factor, traversing all candidate scaling factors to obtain the set of decoding results for all simulated frames corresponding to each candidate scaling factor.

[0095] Step S40105: Based on the simulated decoding and the actual transmitted bits, obtain the simulated decoding accuracy corresponding to the candidate scaling factor.

[0096] In this embodiment, for each frame, the decoded output bit sequence is compared bit-by-bit with the actual transmitted bit sequence. The number of bits that are different is counted as the error bit count for that frame. The error bit counts of all simulated frames are summed to obtain the total error bit count corresponding to the current candidate scaling factor. The total error bit count is divided by the total number of bits in all simulated frames (the total number of bits equals the number of simulated frames multiplied by the number of information bits per frame) to obtain the bit error rate corresponding to the candidate scaling factor. The bit error rate is used as a quantitative indicator of the simulation decoding accuracy. A lower bit error rate indicates that the scaling factor can achieve better decoding performance after calibration. For each candidate scaling factor, the above statistical process is performed independently to obtain the bit error rate value corresponding to each candidate scaling factor.

[0097] Step S40106: Calculate the average decoding accuracy of each candidate scaling factor, and select the candidate scaling factor with the highest average decoding accuracy as the initial value of the scaling factor.

[0098] In this embodiment, based on the bit error rate (BER) results corresponding to each candidate scaling factor, a summary comparison is performed for each candidate scaling factor. For each candidate scaling factor, its corresponding BER is the global average value obtained from all signal-to-noise ratio (SNR) points and all simulation frames. This value directly reflects the comprehensive impact of the scaling factor on decoding performance after calibration. Among all candidate scaling factors, the one with the lowest BER is selected as the optimal value. Since a lower BER indicates a higher decoding accuracy, this optimal value is the scaling factor with the highest average decoding accuracy. This optimal value is determined as the initial value of the scaling factor and written into the non-volatile memory unit of the receiving device for use in the online inference stage. For example, in the simulation calibration of near-Earth satellite scenarios, after the above grid search and statistical comparison, the initial value of the scaling factor is usually around 1.2. This value makes the magnitude distribution of the corrected log-likelihood ratio vector match the input expectation of the LDPC decoder, thereby obtaining the optimal initial decoding performance. This offline calibration process is completed once before system deployment. Once the calibration result is determined, it can be used directly during online inference without repeated calculation. If the average decoding accuracy corresponding to multiple candidate scaling factors is the same, the candidate scaling factor with the smallest value is selected as the initial value.

[0099] Finally, the magnitude of the corrected log-likelihood ratio vector is scaled and adjusted based on the initial value of the scaling factor, and the magnitude distribution of the corrected log-likelihood ratio is calibrated to the optimal input range of the LDPC decoder, thus obtaining the initially adjusted log-likelihood ratio vector, which provides more adaptable soft information input for subsequent channel decoding.

[0100] Step S402: Based on the initial adjusted log-likelihood ratio vector, perform decoding processing and cyclic redundancy check, and then use a sliding window to calculate the decoding accuracy.

[0101] In this embodiment, the initially adjusted log-likelihood ratio vector is first input into the LDPC decoder for decoding. The decoder uses the belief propagation algorithm with a fixed number of iterations of 5, outputting the decoded information bit sequence. Then, a cyclic redundancy check (CRC) is performed on the decoded information bit sequence. This check uses the same generator polynomial as the transmitter, calculating the check bits and comparing them with the received check bits to determine if the frame is decoded correctly. The CRC result of each frame is used as a decoding correctness flag and stored in a fixed-length sliding window. The sliding window length is pre-configured according to service requirements and channel change rate; in this embodiment, the window length W = 50 frames. The window uses a first-in-first-out (FIFO) queue structure. After processing each frame, the CRC result of the current frame (marked as 1 for success and 0 for failure) is enqueued, while the earliest frame result at the head of the queue is removed from the queue. The decoding accuracy within the window is obtained by dividing the sum of the verification results of all frames in the queue by the window length, which is the percentage of frames that pass verification within the window. This accuracy reflects the decoding performance of the most recent W frames and serves as the basis for subsequent steps to determine whether the scaling factor needs to be adjusted. The window statistics process runs continuously online, and the statistics are updated once after each frame is decoded.

[0102] Step S403: When it is determined that the decoding accuracy is lower than the preset benchmark accuracy, the difference between the decoding accuracy and the benchmark accuracy is calculated.

[0103] In this embodiment, a pre-set baseline accuracy rate is first obtained from the system configuration parameters. This baseline accuracy rate is set according to the service quality requirements, and in this embodiment, it is set to 95%. Simultaneously, a pre-configured dead zone threshold is obtained, which is set to 2% in this embodiment. This value, verified through simulation, strikes a balance between preventing frequent adjustments and timely response performance degradation. The actual decoding accuracy rate obtained from the sliding window statistics is recorded as the current accuracy rate. The current accuracy rate is compared with the baseline accuracy rate minus the dead zone threshold, i.e., it is determined whether the current accuracy rate is less than 95% minus 2% (i.e., 93%). If the current accuracy rate is greater than or equal to 93%, the decoding performance is considered to meet the requirements, no adjustment is triggered, and the current scaling factor is used directly to continue processing subsequent frames. If the current accuracy rate is less than 93%, the decoding performance is determined to have degraded, and the scaling factor needs to be adjusted. At this time, the difference between the current accuracy rate and the baseline accuracy rate is calculated, i.e., 95% minus the current accuracy rate, resulting in a positive difference value. This difference value reflects the gap between the current performance and the target performance; the larger the difference, the more severe the performance degradation, and the larger the subsequent adjustment range.

[0104] Step S404: Adjust the initial value of the scaling factor based on the difference to obtain the updated scaling factor.

[0105] Adjusting the initial value of the scaling factor based on the difference to obtain the updated scaling factor specifically includes the following sub-steps: Step S40401: Obtain the deviation and adjustment data pairs through simulation.

[0106] In this embodiment, a communication link model consistent with the target scenario is first constructed in an offline simulation environment, including the processing flow of the transmitter, channel, and receiver. A set of typical deviation values ​​are set as simulation targets. In this embodiment, deviation values ​​of 2%, 5%, 8%, 10%, 12%, 15%, 18%, and 20% are selected, covering the full range from slight performance degradation to severe performance deterioration. For each preset deviation value, the additive noise power or channel fading depth of the receiver is first adjusted so that the decoding accuracy of the receiver using the initial value of the scaling factor is exactly lower than the baseline accuracy of 95% by the corresponding deviation value. Then, under this fixed deviation condition, the scaling factor is used as the variable to be optimized, and a fine-grid search is performed in the range of 0.5 to 2.0 with a step size of 0.02. For each candidate scaling factor, it is applied to the calibration of the corrected log-likelihood ratio vector. The calibrated vector is sent to the decoder for decoding, and the bit error rate under the scaling factor is calculated. The change in the scaling factor that minimizes the bit error rate relative to the current scaling factor is recorded as the optimal adjustment value corresponding to the deviation value. For example, when the deviation value is 10%, the search finds that increasing the current scaling factor by 0.12 can restore the decoding performance to its optimal state, so the recorded data pair is (10%, 0.12). The above simulation search process is repeated for all preset deviation values ​​to obtain a set of data pairs of deviation values ​​and optimal adjustment values.

[0107] Step S40402: Fit the deviation and adjustment data pair to obtain a deviation adjustment relationship model.

[0108] In this embodiment, the deviation is used as the independent variable and the optimal adjustment value as the dependent variable. Data points are plotted in a two-dimensional coordinate system, and their distribution characteristics are observed. In the low deviation region (deviation less than 5%), the adjustment value increases approximately linearly with the deviation, with a relatively gentle slope. In the high deviation region (deviation greater than 10%), the adjustment value increases rapidly with the deviation, and the slope increases. To accurately characterize this nonlinear relationship, a piecewise linear fitting strategy is adopted: the deviation range is divided into three sub-intervals: 0% to 5%, 5% to 10%, and 10% to 20%. Linear regression fitting is performed within each sub-interval to obtain three linear sub-models. Taking an 8% deviation as an example, it falls within the 5% to 10% interval; substituting this into the linear formula for that interval allows for the calculation of the corresponding adjustment value. After fitting, the endpoint values, slopes, and intercept parameters of each interval of the piecewise linear model are stored in the receiving device, forming a deviation adjustment relationship model. This model can be used directly in the online adjustment phase; the current deviation value is input, and the corresponding scaling factor adjustment value is output, eliminating the need for re-simulation or iterative search, ensuring the real-time nature of the adjustment process.

[0109] Step S40403: Input the difference into the deviation adjustment relationship model to obtain the updated scaling factor.

[0110] In this embodiment, the calculated difference is used as input and substituted into the deviation adjustment relationship model. Based on the deviation interval where the difference lies, the corresponding piecewise linear sub-model is selected to calculate the adjustment amount of the scaling factor. A positive adjustment amount indicates that the scaling factor needs to be increased, and a negative value indicates that the scaling factor needs to be decreased. Then, the currently used scaling factor is added to the adjustment amount to obtain the updated scaling factor. To avoid invalid values ​​(such as negative numbers or excessively large values) in the scaling factor, the effective range of the scaling factor is set to [value missing]. If the updated value exceeds this range, the nearest boundary value will be used.

[0111] Step S405: Calculate the final log-likelihood ratio vector based on the updated scaling factor.

[0112] In this embodiment, after updating the scaling factor, the updated scaling factor is applied to the calibration process of the current frame. This involves dividing each element of the corrected log-likelihood ratio vector by the scaling factor to obtain the calibrated log-likelihood ratio vector. This vector is then input into the decoder for decoding, and the decoding accuracy of the frame is calculated. If the decoding accuracy of the frame is still lower than the baseline accuracy minus the dead zone threshold, the scaling factor is iterated and adjusted repeatedly. The updated scaling factor is applied to the calibration process of the next frame until one of the following stopping conditions is met: the decoding accuracy reaches or exceeds the baseline accuracy minus the dead zone threshold; or the number of consecutive adjustments reaches a preset upper limit (3 times in this embodiment). If the process stops due to reaching the upper limit, the current scaling factor is recorded and used in subsequent frames. Simultaneously, the system marks the channel condition as abnormal during this period, which may trigger an alarm or adjust the baseline parameters. This iterative mechanism ensures that the scaling factor can be continuously optimized when performance degrades until satisfactory performance is restored or it is confirmed that recovery is impossible.

[0113] Step S5: Input the final log-likelihood ratio vector into the channel decoder for decoding.

[0114] In this embodiment, the decoder uses the belief propagation algorithm, with a maximum number of iterations of 5. By iteratively updating the messages between the variable nodes and the check nodes, it gradually converges to the valid codeword. During the decoding process, if the check matrix satisfies all check equations within 5 iterations, the iteration is terminated early and the decoded information bit sequence is output. If the maximum number of iterations is reached but the check equations are not satisfied, the hard decision bit sequence obtained in the current iteration is output. The decoder outputs the information bit sequence, which is the final decoding result.

[0115] Step S6: Based on the decoding process, information bits are obtained, and cyclic redundancy check is performed on the information bits to obtain a high-confidence decoding result.

[0116] In this embodiment, after obtaining the information bit sequence from the decoder output, a cyclic redundancy check (CRC) is performed. This check uses the same generator polynomial as the transmitter, recalculating the check bits for the received information bits and comparing them with the check bits carried in the decoding result. If they match, the frame is determined to be correctly decoded and marked as a high-confidence decoding result. If they do not match, the decoding is determined to be incorrect. To further improve the reliability of false labels, an additional confidence condition is added: the actual number of iterations used by the decoder is recorded. Only when the number of iterations is less than a preset threshold (set to 4 in this embodiment) and the CRC passes is the decoding result confirmed as a high-confidence sample; otherwise, it is not used for subsequent incremental training. This threshold of 4 was determined through simulation experiments: when the decoder converges within 4 iterations, its accuracy is usually higher than 95%, and the false detection probability is low. Through the dual screening of CRC and the number of iterations, the samples used for model updates are ensured to have high confidence, reducing the negative impact of erroneous labels on incremental training.

[0117] Step S7: Associate the high-confidence decoding result with the corresponding initial log-likelihood ratio vector and the channel-related feature matrix to form incremental training samples.

[0118] The process of associating the high-confidence decoding result with the corresponding initial log-likelihood ratio vector and the channel-related feature matrix to form incremental training samples specifically includes the following sub-steps: Step S701: Obtain the pseudo-tag log-likelihood ratio based on the high-confidence decoding result and the received signal.

[0119] In this embodiment, for the confirmed high-confidence decoding result, it is first converted into a reconstructed transmit symbol sequence according to the QPSK modulation mapping method, denoted as... ,in This indicates the symbol position. This is then used to reconstruct the transmitted symbol and received signal. Channel estimation value ( , , To estimate the phase of the channel, The imaginary unit satisfies ) and noise variance Calculate the pseudo-tag log-likelihood ratio for each bit.

[0120] The log-likelihood ratio of pseudo-labels satisfies the following formula: in The two bits representing this symbol, Indicates complex conjugation. This represents taking the real part. This formula is equivalent to projecting the error between the reconstructed symbol and the equalized symbol onto the symbol direction, reflecting the degree of matching between the received signal and the high-confidence decoding result. This formula is applied to all... symbols, resulting in a length of The pseudo-label is a log-likelihood ratio vector. Ideally, this vector should be close to the corrected log-likelihood ratio vector from step S307, but there is a difference between the two due to channel noise and model error. By using this pseudo-label as a training target, the model can learn a mapping relationship that is closer to the real channel characteristics.

[0121] Step S702: Smooth the log-likelihood ratio of the pseudo-label to obtain the smoothed pseudo-label.

[0122] In this embodiment, to reduce label noise caused by false detections in cyclic redundancy check (CR) under low signal-to-noise ratio (SNR) conditions, the pseudo-label log-likelihood ratio is smoothed. The smoothing coefficient is dynamically determined based on the SNR estimate of the current frame: 0.02 for SNR above 10dB, 0.05 for SNR between 0dB and 10dB, and 0.1 for SNR below 0dB. This smoothing method uses multiplicative scaling, which linearly reduces the confidence of pseudo-labels and is simple to implement, facilitating embedded deployment. The smoothed pseudo-label vector is obtained by multiplying the pseudo-label log-likelihood ratio by (1 minus the smoothing coefficient). This processing appropriately reduces the confidence of pseudo-labels, avoiding overfitting of the model to potentially incorrect labels, enhancing the robustness of incremental training. The smoothed pseudo-label vector has the same dimension and numerical range as the original initial log-likelihood ratio vector and can be directly used as a training target.

[0123] Step S703: Associate the smoothed pseudo-label with the corresponding initial log-likelihood ratio vector and the channel-related feature matrix to form labeled training samples.

[0124] In this embodiment, for each frame of samples identified as having high confidence and for which smoothed pseudo-labels are generated, the initial log-likelihood ratio vector, the channel-related feature matrix, and the smoothed pseudo-label vector are stored together. Specifically, the initial log-likelihood ratio vector and the channel-related feature matrix are used as input features, and the smoothed pseudo-label vector is used as the corresponding supervision target; these three together form a labeled training sample. Each sample is independently stored in the non-volatile storage buffer of the receiving device, with a preset buffer capacity of 500 samples, managed using a first-in-first-out (FIFO) queue. When the buffer is full, the newest sample overwrites the oldest sample, ensuring that the buffer always stores the latest high-confidence samples. This sample set will serve as the data source for subsequent incremental training, used to update the log-likelihood ratio correction model, enabling the model to continuously adapt to the slow changes in channel characteristics.

[0125] Step S8: Update the log-likelihood ratio correction model using the incremental training samples.

[0126] In this embodiment, an asynchronous background update mechanism is adopted to ensure that model updates do not affect real-time decoding processing. The system maintains a low-priority background process that periodically checks the cache. When the number of cached samples reaches a preset threshold (500 in this embodiment) or a timer expires (e.g., every hour), the background process starts offline incremental training. During training, the labeled training samples in the cache are merged with a small number of original benchmark samples. It is not necessary to load the full amount of original data. The training workload is the same as training new samples alone. The same loss function (e.g., cross-entropy loss) and optimizer (Adam, initial learning rate 1e-4) as the initial training are used, but the learning rate is reduced to one-tenth of the initial value. The number of training rounds is set to 5 to 10, the batch size is set to 32, and the currently used log-likelihood ratio correction model is fine-tuned. After training is completed, the new model is temporarily stored in the backup area. During system idle periods (e.g., gaps when no data is received), atomic replacement is performed: a double buffering technique is used to point the online model pointer to the new model, while the old model is retained until it is no longer referenced and then released. This mechanism ensures that the switching process has no impact on real-time decoding. Through this mechanism, the model can continuously adapt to the long-term evolution of channel characteristics and maintain the stability of decoding performance.

[0127] like Figure 2 As shown, in another aspect, the present invention also provides a channel decoding system based on a dynamic channel environment, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the relevant steps of a relevant embodiment of the channel decoding method based on a dynamic channel environment of the present invention.

[0128] This invention provides a channel decoding system based on a dynamic channel environment. The functional components can be integrated into a single processing unit, or each component can exist independently, or two or more components can be integrated into a single unit. The integrated components can be implemented in hardware or as software functions.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for channel decoding based on dynamic channel environment, characterized by, The method includes: The received signal is acquired, and the received signal is synchronized, channel estimated, and demodulated to obtain an initial log-likelihood ratio vector and a channel correlation feature matrix. The initial log-likelihood ratio vector and the channel-related feature matrix are concatenated and normalized to obtain a normalized feature matrix; Construct a log-likelihood ratio correction model, input the normalized feature matrix into the log-likelihood ratio correction model, and obtain the corrected log-likelihood ratio vector; Obtain the initial value of the pre-set scaling factor, and perform calibration processing on the corrected log-likelihood ratio vector to obtain the final log-likelihood ratio vector; The final log-likelihood ratio vector is input into the channel decoder for decoding.

2. The method of claim 1, wherein, The process of synchronizing, estimating, and demodulating the received signal to obtain the initial log-likelihood ratio vector and the channel correlation feature matrix includes: The received signal is subjected to carrier synchronization and symbol synchronization processing to obtain a synchronized received signal; Channel estimation is performed on the synchronized received signal to obtain the channel estimation amplitude, channel estimation phase, and signal-to-noise ratio estimation values; The synchronized received signal is demodulated, and the initial log-likelihood ratio vector is calculated by combining the channel estimation amplitude, the channel estimation phase, and the signal-to-noise ratio estimate. Extract the historical log-likelihood ratio vector of adjacent code blocks of the synchronized received signal; The historical log-likelihood ratio vector, the channel estimation amplitude, the channel estimation phase, and the signal-to-noise ratio estimate are concatenated by channel to form a multi-dimensional feature matrix, thus obtaining the channel-related feature matrix.

3. The method of claim 1, wherein, The process of concatenating and normalizing the initial log-likelihood ratio vector and the channel-related feature matrix to obtain the normalized feature matrix includes: The initial log-likelihood ratio vector and the channel-related feature matrix are concatenated by channel to form a multi-channel feature matrix; Obtain the pre-calculated global mean and global standard deviation; The multi-channel feature matrix is ​​standardized using the global mean and global standard deviation to obtain a normalized feature matrix.

4. The method of claim 1, wherein, The construction of the log-likelihood ratio correction model includes: Obtain the maximum Doppler frequency shift and symbol period of the target scene; The coherence time is calculated based on the maximum Doppler frequency shift. The number of symbols contained in the coherence time is calculated based on the symbol period; The total number of dilated convolutional layers is determined based on the number of symbols. The dilation coefficient of each of the dilated convolutional layers is determined based on the total number of layers and the number of symbols. Construct a dilated convolutional residual module based on the dilated convolutional layer and the dilation coefficient; A log-likelihood ratio correction model is constructed based on the pre-built normalization module and dimension compression module combined with the dilated convolution residual module.

5. The method of claim 1, wherein, The steps of obtaining a pre-set initial value for the scaling factor and calibrating the corrected log-likelihood ratio vector to obtain the final log-likelihood ratio vector include: Obtain a preset initial value for the scaling factor, and perform an initial adjustment on the corrected log-likelihood ratio vector to obtain the initially adjusted log-likelihood ratio vector. After decoding based on the initially adjusted log-likelihood ratio vector and performing cyclic redundancy checks, the decoding accuracy is statistically analyzed using a sliding window. When the decoding accuracy is determined to be lower than a preset benchmark accuracy, the difference between the decoding accuracy and the benchmark accuracy is calculated. The initial value of the scaling factor is adjusted based on the difference to obtain the updated scaling factor; The final log-likelihood ratio vector is calculated based on the updated scaling factor.

6. The method of claim 5, wherein, The process of obtaining the preset initial value of the scaling factor includes: Obtain the channel parameters of the target dynamic channel scenario, and generate a simulated received signal and the corresponding real transmitted bits based on the channel parameters; The simulated received signal is processed using the log-likelihood ratio correction model to obtain the simulated corrected log-likelihood ratio vector. Define a range of candidate scaling factors and use a grid search algorithm to traverse the candidate scaling factors within that range. The simulation-corrected log-likelihood ratio vector is calibrated and decoded using the candidate scaling factors to obtain the simulation decoding. Based on the simulated decoding and the actual transmitted bits, the simulation decoding accuracy corresponding to the candidate scaling factor is obtained; The average decoding accuracy of each candidate scaling factor is calculated, and the candidate scaling factor with the highest average decoding accuracy is selected as the initial value of the scaling factor.

7. The method of claim 5, wherein the step of decoding the channel comprises the step of: The step of adjusting the initial value of the scaling factor based on the difference to obtain the updated scaling factor includes: ​ Data pairs of deviation and adjustment values ​​are obtained through simulation. A deviation adjustment relationship model is obtained by fitting the deviation and adjustment data pairs. The difference is input into the deviation adjustment relationship model to obtain the updated scaling factor.

8. The method of claim 1, wherein, The method further includes: Information bits are obtained based on the decoding process, and high-confidence decoding results are obtained by performing cyclic redundancy check on the information bits. The high-confidence decoding results are correlated with the corresponding initial log-likelihood ratio vector and the channel correlation feature matrix to form incremental training samples; The log-likelihood ratio correction model is updated using the incremental training samples.

9. A channel decoding method based on a dynamic channel environment according to claim 8, characterized in that, The step of associating the high-confidence decoding result with the corresponding initial log-likelihood ratio vector and the channel correlation feature matrix to form incremental training samples includes: The pseudo-tag log-likelihood ratio is obtained based on the high-confidence decoding result and the received signal; The log-likelihood ratio of the pseudo-labels is smoothed to obtain smoothed pseudo-labels; The smoothed pseudo-labels are associated with the corresponding initial log-likelihood ratio vector and the channel-related feature matrix to form labeled training samples.

10. A channel decoding system based on a dynamic channel environment, characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute a channel decoding method based on a dynamic channel environment as described in any one of claims 1 to 9.

Images