Lightweight dual-binary Turbo decoding method and system based on adaptive channel perception

Abstract

The invention relates to a lightweight dual-binary Turbo decoding method and system based on adaptive channel perception. The method comprises the following steps: data acquisition and preprocessing; the decoding model comprises a dynamic channel sensing module, a bidirectional GRU network, a multi-branch decoding network, a residual error correction unit and an iteration termination controller. Comprising the following steps: extracting kurtosis, instantaneous signal-to-noise ratio and zero crossing point rate characteristics of a received signal by adopting a sliding window; noise type classification is carried out through a bidirectional GRU network; dynamically activating a decoding branch in the multi-branch decoding network with the corresponding complexity, and fusing a traditional Max-Log-MAP algorithm output and a neural network prediction value; dynamic optimization of the decoding process is realized through an iteration termination controller based on confidence coefficient change; and performing end-to-end training on the decoding model, and performing hard decision on the final LLR value through a sigmoid function to output a decoding result. According to the invention, a closed-loop system of channel perception-decoding decision-resource optimization is constructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of channel coding and signal processing in wireless communication systems, and particularly relates to a lightweight duobinary Turbo decoding method and system based on adaptive channel perception. Background Art

[0002] Duobinary Turbo codes are widely used in communication standards such as LTE and WiMAX due to their excellent error correction performance and moderate decoding complexity. However, traditional decoding methods have the following technical bottlenecks:

[0003] (1) Limitations of static computing architecture

[0004] Existing algorithms such as Log-MAP and Max-Log-MAP use a fixed number of iterative decoding. When channel conditions deteriorate (such as multipath fading and sudden interference), the number of iterations must be increased to maintain the bit error rate, resulting in a linear increase in computing resources, making it difficult to meet the requirements of URLLC (Ultra-Reliable Low Latency Communication) scenarios.

[0005] (2) Insufficient environmental adaptability

[0006] Traditional methods assume that channel noise follows a steady-state Gaussian distribution. However, real wireless environments often contain non-Gaussian noise (such as radar pulse interference and electromagnetic noise from industrial equipment) and non-stationary characteristics (such as fast-varying Rayleigh fading). In such scenarios, the performance of fixed-parameter decoders degrades significantly.

[0007] (3) Deep Learning Model Deployment Barriers

[0008] Although the neural network decoders based on RNN and Transformer proposed in recent years can improve noise resistance, their parameter count usually reaches millions, resulting in excessive memory usage and energy consumption, making them difficult to deploy on terminal devices. Summary of the Invention

[0009] In view of the shortcomings of the existing technology, the present invention provides a lightweight dual binary Turbo decoding method based on adaptive channel perception;

[0010] This method combines dynamic channel state estimation, a lightweight deep learning model, and a traditional belief propagation algorithm to achieve low-complexity, high-reliability decoding in highly dynamic environments. This method is suitable for scenarios requiring both low latency and high energy efficiency, such as 5G / 6G mobile communications, satellite communications, and vehicle-to-everything (V2X).

[0011] The present invention also provides a lightweight dual binary Turbo decoding system based on adaptive channel perception.

[0012] Explanation of terms:

[0013] The traditional Max-Log-MAP algorithm is one of the core algorithms of Turbo decoding. It reduces computational complexity by approximating the maximum likelihood path metric. Its core formula is:

[0014]

[0015] in is the state transition metric, is the forward state transfer metric, is the backward state transition metric.

[0016] The technical solution of the present invention is:

[0017] A lightweight duobinary turbo decoding method based on adaptive channel sensing, comprising:

[0018] Data acquisition and preprocessing; data refers to the demodulated soft decision sequence; data preprocessing refers to converting the demodulated soft decision sequence into amplitude and phase information including symbol level;

[0019] The pre-processed data is fed into the trained decoding model to implement duobinary turbo decoding;

[0020] The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including:

[0021] In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type;

[0022] Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit;

[0023] Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes;

[0024] The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

[0025] Preferably, according to the present invention, in the dynamic channel sensing module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio and zero-crossing rate characteristics of the received signal; comprising:

[0026] Perform windowing on the demodulated soft decision sequence and calculate the kurtosis, instantaneous signal-to-noise ratio and zero-crossing rate in each window;

[0027] The kurtosis calculation formula is as follows:

[0028]

[0029] Among them, K is the kurtosis, x i is the i-th sampling value of the received signal, μ is the mean of the received signal in the sliding window, σ is the standard deviation of the received signal in the sliding window, and N is the length of the sliding window;

[0030] The instantaneous signal-to-noise ratio is calculated as follows:

[0031]

[0032] Among them, SNR inst is the instantaneous signal-to-noise ratio, y is the received signal, n is the estimated noise, ||y|| 2 is the received signal power, ||n|| 2 is the noise power;

[0033] The formula for calculating the zero-crossing rate is as follows:

[0034]

[0035] The zero crossing rate ZCR is used to detect phase jumps and identify multipath effects; i is the i-th sampling value of the original transmitted signal, y i+1 Indicates adjacent sampling points;

[0036] sign() is the symbol function defined as:

[0037]

[0038] According to a preferred embodiment of the present invention, noise type classification is performed by a bidirectional GRU network; comprising:

[0039] The feature vector f = [K, SNR inst ,ZCR] input gated recurrent unit, output noise type probability distribution p(noise type|f):

[0040] p(noise type|f)=Softmax(W h GRU(f)+b h );

[0041] Among them, GRU is a gated recurrent unit, the input is a feature vector sequence, and the output is a hidden state, W h is the weight matrix of the fully connected layer, b h is the bias vector of the fully connected layer. The noise types include Gaussian, Rayleigh, and burst. h is the hidden state of GRU for the input sequence f.

[0042] Preferably, according to the present invention, the multi-branch decoding network includes a separable convolution module, the separable convolution module includes a depth convolution layer and a point convolution layer, and the depth convolution layer and the point convolution layer process data in the spatial and channel dimensions respectively;

[0043] Use deep convolutional layers to extract spatial features and perform channel fusion through point convolutional layers;

[0044] The depth convolution layer uses a 3×1 kernel for spatial feature extraction, and each input channel is convolved independently, with a parameter quantity P depthwise for:

[0045] P depthwise =K h ×K w ×C in ;

[0046] Among them, K h is the height of the convolution kernel, K w Indicates the width of the convolution kernel, C in Indicates the number of input channels;

[0047] The point convolution layer uses a 1×1 convolution kernel to perform channel fusion on the spatial features of the depth convolution extraction layer; the parameter quantity P point for:

[0048] P point =C in ×C out ;

[0049] The total number of parameters is reduced to that of standard convolution times; among them, K h ×K w is the convolution kernel size, C in is the number of input channels, C out is the number of output channels.

[0050] Preferably, according to the present invention, the decoding branches in the multi-branch decoding network of corresponding complexity are dynamically activated according to the noise type classification result; comprising:

[0051] In the Gaussian noise scenario, the low-complexity branch is activated. The low-complexity branch includes three separable convolutional layers and an average pooling layer. Gaussian noise refers to SNR>8dB.

[0052] In Rayleigh fading scenarios, the medium-complexity branch is activated. The medium-complexity branch includes 5 separable convolutional layers and skip connections. Rayleigh fading refers to 4dB≤SNR≤8dB.

[0053] In the burst interference scenario, the high-complexity branch is activated. The high-complexity branch includes 7 separable convolutional layers and a channel attention mechanism. Burst interference refers to SNR < 4dB.

[0054] According to a preferred embodiment of the present invention, the output of the traditional Max-Log-MAP algorithm is fused with the prediction value of the neural network through a residual correction unit; comprising:

[0055] The traditional Max-Log-MAP algorithm outputs the initial LLR value LLR(x), which is as follows:

[0056]

[0057] Where y is the received signal and P(y|x) is the conditional probability of receiving the signal given x.

[0058] Neural network prediction residual correction term:

[0059] LLR final =LLR MAP +α·ΔLLR;

[0060] Among them, LLR final is the final log-likelihood ratio, ΔLLR is the LLR residual correction term predicted by the neural network, and LLR MAP It means that the LLR value α output by the traditional Max-Log-MAP algorithm is the adaptive weight;

[0061] α is calculated by the following formula:

[0062] α=σ(W σ ·[SNR inst ,K]+b α );

[0063] σ is the Sigmoid function, which constrains α to be in the interval [0.3, 1.0]; W σ and b α is the weight matrix and bias term, which are learned through training.

[0064] Preferably, according to the present invention, dynamic optimization of the decoding process is achieved by an iterative termination controller based on changes in confidence, comprising:

[0065] Design a dual-threshold early stopping strategy as follows:

[0066]

[0067] Among them, T is the total number of bits of a frame of data, δ is the confidence change threshold, softmax(LLR k ) indicates that the LLR value is normalized and converted into a probability distribution. represents the log-likelihood ratio of the t-th bit during the k-th iteration; represents the log-likelihood ratio of the t-th bit during the k-1-th iteration.

[0068] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned lightweight dual binary Turbo decoding method based on adaptive channel sensing when executing the computer program.

[0069] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned lightweight dual binary Turbo decoding method based on adaptive channel sensing.

[0070] A lightweight duobinary turbo decoding system based on adaptive channel sensing includes:

[0071] The data acquisition and preprocessing module is configured as follows: data acquisition and preprocessing; data refers to a demodulated soft decision sequence; data preprocessing refers to converting the demodulated soft decision sequence into amplitude and phase information including symbol level;

[0072] The duobinary turbo decoding module is configured to: input the preprocessed data into the trained decoding model to implement duobinary turbo decoding;

[0073] The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including:

[0074] In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type;

[0075] Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit;

[0076] Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes;

[0077] The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

[0078] The beneficial effects of the present invention are:

[0079] 1. Improved environmental adaptability: This invention uses a dynamic channel sensing module to identify noise types in real time, adapting to changing wireless environments. This eliminates the need to rely on fixed noise models, improving decoding reliability.

[0080] 2. Reduced computational complexity: Dynamically selects decoding branches of varying complexity based on noise type, optimizing computing resource usage. This is particularly useful for devices with limited computing resources, improving decoding efficiency.

[0081] 3. Optimized Bit Error Performance: The residual correction unit combines traditional algorithms with neural network prediction to improve decoding accuracy. The iterative termination controller reduces redundant calculations and further optimizes the bit error rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is the overall architecture diagram of a lightweight duobinary Turbo decoding method based on adaptive channel perception;

[0083] Figure 2 A schematic diagram of the process of dynamically activating the dynamic decoding branch;

[0084] Figure 3 Schematic diagram of the residual correction unit structure;

[0085] Figure 4 Block diagram of the separable convolution module. DETAILED DESCRIPTION

[0086] The present invention will be further defined below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0087] Example 1

[0088] A lightweight dual binary turbo decoding method based on adaptive channel sensing, such as Figure 1 As shown, including:

[0089] Data acquisition and preprocessing; data refers to the soft decision sequence after demodulation at the receiving end; data preprocessing refers to converting the soft decision sequence after demodulation at the receiving end into amplitude and phase information including symbol level;

[0090] The pre-processed data is fed into the trained decoding model to implement duobinary turbo decoding;

[0091] The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including:

[0092] In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type;

[0093] Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit;

[0094] Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes;

[0095] The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

[0096] The present invention combines adaptive channel perception with a deep learning model to construct a closed-loop system of channel perception-decoding decision-resource optimization.

[0097] Example 2

[0098] The lightweight duobinary Turbo decoding method based on adaptive channel sensing described in Example 1 is different in that:

[0099] In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio (SNR), and zero-crossing rate (ZCR) features of the received signal; these features are then input into a bidirectional RNN (BRNN) network for time series information processing, ultimately outputting a probability distribution of the noise type. This includes:

[0100] The demodulated soft decision sequence is windowed (window length 5ms, overlap rate 50%), and the kurtosis, instantaneous signal-to-noise ratio (SNR) and zero-crossing rate (ZCR) are calculated in each window.

[0101] The kurtosis calculation formula is as follows:

[0102]

[0103] Among them, K is the kurtosis, which is used to quantify the "peaked and thick-tailed" characteristics of the noise distribution and distinguish Gaussian noise (K≈0) from burst interference (K>3). i is the i-th sampling value of the received signal, μ is the mean of the received signal in the sliding window, σ is the standard deviation of the received signal in the sliding window, and N is the length of the sliding window;

[0104] The instantaneous signal-to-noise ratio is calculated as follows:

[0105]

[0106] Among them, SNR inst is the instantaneous signal-to-noise ratio, y is the received signal, n is the estimated noise, ||y|| 2 is the received signal power, ||n|| 2 is the noise power;

[0107] The formula for calculating the zero-crossing rate is as follows:

[0108]

[0109] The zero crossing rate ZCR is used to detect phase jumps and identify multipath effects; i is the i-th sampling value of the original transmitted signal, y i+1 Indicates adjacent sampling points;

[0110] sign() is the symbol function defined as:

[0111]

[0112] Noise type classification is performed using a bidirectional GRU network; including:

[0113] The feature vector f = [K, SNR inst ,ZCR] input gated recurrent unit (GRU), output noise type probability distribution p(noise type|f):

[0114] p(noise type|f)=Softmax(W h GRU(f)+b h );

[0115] Among them, GRU is a gated recurrent unit, the input is a feature vector sequence, and the output is a hidden state, W h is the weight matrix of the fully connected layer, b h is the bias vector of the fully connected layer. The noise types include Gaussian, Rayleigh, and burst. h is the hidden state of the GRU for the input sequence f. It contains important information in the sequence and plays a key role in noise classification.

[0116] The bidirectional GRU network architecture builds on the traditional GRU by combining forward and reverse GRU layers to better capture temporal information in sequences. The forward GRU processes the input sequence from left to right, sequentially reading the feature vector sequence and outputting the corresponding hidden states. The reverse GRU processes the input sequence from right to left, sequentially reading the feature vector sequence and outputting the corresponding hidden states.

[0117] In the forward GRU, each moment of the input sequence generates a hidden state n fozward , which represents the contextual information of the sequence from left to right. In the reverse GRU, each moment of the input sequence generates a hidden state h backward , indicating the contextual information of the sequence from right to left.

[0118] The bidirectional GRU concatenates the hidden states of these two directions to obtain the final hidden state:

[0119] h {i =[h fozward ,h backward ];

[0120] Among them, h bi represents the hidden state of the bidirectional GRU at each moment, which contains richer contextual information in the sequence.

[0121] Hidden state h bi Will be passed to the fully connected layer (i.e. W h ·h bi +b h ), and then the probability distribution of noise type is obtained through the Softmax activation function.

[0122] Expand the traditional Max-Log-MAP algorithm into a parameterized deep neural network and construct a multi-branch decoding network with separable convolution and attention mechanisms;

[0123] The multi-branch decoding network includes separable convolution modules, such as Figure 4 As shown in the figure, the separable convolution module includes a depth convolution layer and a point convolution layer, which process data in the spatial and channel dimensions respectively;

[0124] Deep convolutional layers are used to extract spatial features, and point convolutional layers are used for channel fusion, significantly reducing the amount of computation.

[0125] The depthwise convolution layer (Depthwise Conv) uses a 3×1 kernel for spatial feature extraction, and each input channel is convolved independently, with a parameter quantity P depthwise for:

[0126] P depthwise =K h ×K w ×C in ;

[0127] Among them, K h is the height of the convolution kernel. In the present invention, K h is 3; K w Represents the width of the convolution kernel. In this invention, K w is 1; C in Indicates the number of input channels;

[0128] The pointwise convolution layer uses a 1×1 convolution kernel to perform channel fusion on the spatial features of the depth convolution extraction layer; it is usually used to integrate features in the channel dimension so that information between different channels can be transferred to each other. point for:

[0129] P point =C in ×C out ;

[0130] The total number of parameters is reduced to that of standard convolution times; among them, K h ×K w is the convolution kernel size, C in is the number of input channels, C out is the number of output channels.

[0131] According to the noise type classification result, the decoding branches in the multi-branch decoding network with corresponding complexity are dynamically activated; Figure 2 As shown, including:

[0132] Based on the predicted noise type, the system can select the appropriate decoding branch. For example, under high SNR conditions, a low-complexity branch is selected, while under low SNR conditions, a high-complexity branch containing an RNN is selected to handle burst interference.

[0133] In Gaussian noise scenarios, the low-complexity branch is activated. The low-complexity branch consists of three separable convolutional layers and an average pooling layer. Gaussian noise refers to an SNR greater than 8dB. The separable convolutional layer extracts spatial features through depthwise convolution, and then performs channel fusion through pointwise convolution, thereby reducing computational complexity. The average pooling layer downsamples the input and calculates the average value of the input region. This helps reduce computational effort and improves network robustness to a certain extent. It is often used to reduce spatial dimensionality while preserving key information and reducing sensitivity to input noise. In low SNR scenarios (such as sudden interference), a high-complexity network with an RNN and channel-wise attention mechanism is selected.

[0134] In Rayleigh fading scenarios, the medium-complexity branch is activated. It consists of five separable convolutional layers and skip connections. Rayleigh fading refers to a SNR of 4dB ≤ ≤ 8dB. Skip connections allow certain layers in the network to skip layers and connect directly to subsequent layers. Skip connections help alleviate the vanishing gradient problem and improve network training efficiency, especially in deep networks. They also preserve low-level details, facilitating better feature fusion. In medium SNR scenarios (such as Rayleigh fading), a five-layer separable convolutional network is used.

[0135] In burst interference scenarios, the high-complexity branch is activated, consisting of seven separable convolutional layers and a channel attention mechanism. Burst interference refers to an SNR < 4dB. The channel attention mechanism is a mechanism that automatically assigns weights to different channels. Through this mechanism, the network can adaptively focus on more important channel information, thereby suppressing irrelevant features. This is usually achieved through the Squeeze-and-Excitation (SE) module, which performs global information compression on the features of each channel, then learns the weight coefficients for each channel through a fully connected layer, and finally applies these weights to the input features. Under high SNR (such as Gaussian noise), a low-complexity three-layer separable convolutional network is activated.

[0136] This dynamic selection mechanism ensures that the decoding process operates efficiently under different channel conditions.

[0137] The output of the traditional Max-Log-MAP algorithm is fused with the neural network prediction value through the residual correction unit; Figure 3 As shown, including:

[0138] Residual correction: By combining the preliminary LLR values ​​output by the traditional Max-Log-MAP algorithm and the residual correction terms predicted by the deep learning model, the system can accurately calculate the LLR values ​​and perform final decoding.

[0139] Operation: The correction term output by the neural network is weighted and fused with the LLR value output by the traditional algorithm, and the residual correction amount is dynamically adjusted to improve the decoding accuracy.

[0140] The traditional Max-Log-MAP algorithm outputs the initial LLR value LLR(x), which is usually estimated based on the received signal and the predetermined symbol model. The details are as follows:

[0141]

[0142] Where y is the received signal and P(y|x) is the conditional probability of receiving the signal given x.

[0143] The neural network predicts the residual correction term. Its role here is to learn how to correct the LLR output of the traditional algorithm. The network architecture is typically a deep network that uses the input received signal (or intermediate network features) to predict a residual correction term, ΔLLR. This correction term is the deviation between the output of the traditional Max-Log-MAP algorithm and the true log-likelihood ratio. The prediction process is as follows: Through training, the neural network learns how to predict the correction term based on the input signal (such as SNR and received symbols), further adjusting the accuracy of the LLR value. The neural network can be a model containing multiple convolutional layers, fully connected layers, and other structures. It typically learns effective features in the context of the input signal.

[0144] LLR final =LLR MAP +α·ΔLLR;

[0145] Among them, LLR final is the final log-likelihood ratio, ΔLLR is the LLR residual correction term predicted by the neural network, and LLR MAP It means that the LLR value α output by the traditional Max-Log-MAP algorithm is the adaptive weight;

[0146] α is calculated by the following formula:

[0147] α=σ(W σ ·[SNR inst ,K]+b α );

[0148] σ is the Sigmoid function, which constrains α to be in the interval [0.3, 1.0]; W σ and b α is the weight matrix and bias term, which are learned through training.

[0149] Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes; including:

[0150] Automatic iteration termination is achieved through deep learning models, and redundant calculations are avoided and computational complexity is reduced by predicting the gain of each iteration.

[0151] Design a dual-threshold early stopping strategy as follows:

[0152]

[0153] Among them, T is the total number of bits of a frame of data, δ is the confidence change threshold, softmax(LLR k ) indicates that the LLR value is normalized and converted into a probability distribution. represents the log-likelihood ratio (LLR) value of the t-th bit during the k-th iteration; represents the log-likelihood ratio (LLR) value of the t-th bit during the k-1-th iteration.

[0154] By combining RNN with feedback information, we dynamically determine whether to terminate the decoding process early. By adopting a dual-threshold strategy, we control when to stop iteration, avoiding unnecessary calculations and improving efficiency.

[0155] The Adam optimizer is used to train the decoding model end-to-end, and the final LLR value LLR is calculated by the sigmoid function. final Make hard decisions and output decoding results; including:

[0156] Input data: The training data is usually the received signal obtained through channel simulation. After noise perturbation, the signal will contain erroneous bits. The task of the model is to recover the correct bits from these noisy signals.

[0157] Model structure: A recurrent neural network (RNN)-based structure can be used to process time series data. The neural network learns how to decode channel input data (noisy signals).

[0158] Loss function: The loss function used during training is usually the cross-entropy loss function, which measures the difference between the network output (probability distribution) and the actual label.

[0159] Adam optimizer: Adam optimizer is a gradient-based optimization algorithm that can effectively accelerate the training process of deep learning models.

[0160] Example 3

[0161] A computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the lightweight dual binary turbo decoding method based on adaptive channel sensing described in embodiment 1 or 2 are implemented.

[0162] Example 4

[0163] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the lightweight dual binary turbo decoding method based on adaptive channel perception described in embodiment 1 or 2 are implemented.

[0164] Example 5

[0165] A lightweight duobinary turbo decoding system based on adaptive channel sensing includes:

[0166] The data acquisition and preprocessing module is configured as follows: data acquisition and preprocessing; data refers to a soft decision sequence demodulated by a receiving end; data preprocessing refers to converting the soft decision sequence demodulated by the receiving end into amplitude and phase information including symbol level;

[0167] The duobinary turbo decoding module is configured to: input the preprocessed data into the trained decoding model to implement duobinary turbo decoding;

[0168] The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including:

[0169] In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type;

[0170] Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit;

[0171] Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes;

[0172] The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

Claims

1. A lightweight duobinary turbo decoding method based on adaptive channel sensing, characterized in that: include: Data acquisition and preprocessing; Data refers to the soft decision sequence after demodulation; Data preprocessing refers to converting the demodulated soft decision sequence into amplitude and phase information including symbol level; The pre-processed data is fed into the trained decoding model to implement duobinary turbo decoding; The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including: In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type; Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit; Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes; The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

2. The lightweight duobinary turbo decoding method based on adaptive channel perception according to claim 1, characterized in that: In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; this includes: Perform windowing on the demodulated soft decision sequence and calculate the kurtosis, instantaneous signal-to-noise ratio and zero-crossing rate in each window; The kurtosis calculation formula is as follows: Among them, K is the kurtosis, x i is the i-th sampling value of the received signal, μ is the mean of the received signal in the sliding window, σ is the standard deviation of the received signal in the sliding window, and N is the length of the sliding window; The instantaneous signal-to-noise ratio is calculated as follows: Among them, SNR inst is the instantaneous signal-to-noise ratio, y is the received signal, n is the estimated noise, ||y|| 2 is the received signal power, ||n|| 2 is the noise power; The formula for calculating the zero-crossing rate is as follows: The zero crossing rate ZCR is used to detect phase jumps and identify multipath effects; i is the i-th sampling value of the original transmitted signal, y i+1 Indicates adjacent sampling points; sign() is the symbol function defined as:

3. The lightweight duobinary turbo decoding method based on adaptive channel perception according to claim 1, characterized in that: Noise type classification is performed using a bidirectional GRU network; including: The feature vector f = [K, SNR inst ,ZCR] input gated recurrent unit, output noise type probability distribution p(noise type|f): p(noise type|f)=Softmax(W h GRU(f)+b h ); Among them, GRU is a gated recurrent unit, the input is a feature vector sequence, and the output is a hidden state, W h is the weight matrix of the fully connected layer, b h is the bias vector of the fully connected layer. The noise types include Gaussian, Rayleigh, and burst. h is the hidden state of GRU for the input sequence f.

4. The lightweight duobinary turbo decoding method based on adaptive channel perception according to claim 1, characterized in that: The multi-branch decoding network includes a separable convolution module, which includes a depth convolution layer and a point convolution layer. The depth convolution layer and the point convolution layer process data in the spatial and channel dimensions respectively. Use deep convolutional layers to extract spatial features and perform channel fusion through point convolutional layers; The depth convolution layer uses a 3×1 kernel for spatial feature extraction, and each input channel is convolved independently, with a parameter quantity P depthwise for: P depthwise =K h ×K w ×C in ; Among them, K h is the height of the convolution kernel, K w Indicates the width of the convolution kernel, C in Indicates the number of input channels; The point convolution layer uses a 1×1 convolution kernel to perform channel fusion on the spatial features of the depth convolution extraction layer; the parameter quantity P point for: P point =C in ×C out ; The total number of parameters is reduced to that of standard convolution times; among them, L h ×K w is the convolution kernel size, C in is the number of input channels, C out is the number of output channels.

5. The lightweight duobinary turbo decoding method based on adaptive channel perception according to claim 1, characterized in that: Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results; including: In the Gaussian noise scenario, the low-complexity branch is activated. The low-complexity branch includes three separable convolutional layers and an average pooling layer. Gaussian noise refers to SNR>8dB. In Rayleigh fading scenarios, the medium-complexity branch is activated. The medium-complexity branch includes 5 separable convolutional layers and skip connections. Rayleigh fading refers to 4dB≤SNR≤8dB. In the burst interference scenario, the high-complexity branch is activated. The high-complexity branch includes 7 separable convolutional layers and a channel attention mechanism. Burst interference refers to SNR < 4dB.

6. The lightweight duobinary turbo decoding method based on adaptive channel perception according to claim 1, characterized in that: The output of the traditional Max-Log-MAP algorithm is integrated with the prediction value of the neural network through the residual correction unit; include: The traditional Max-Log-MAP algorithm outputs the initial LLR value LLR(x), which is as follows: Where y is the received signal and P(y|x) is the conditional probability of receiving the signal given x. Neural network prediction residual correction term: LLR final =LLR mAP +α·ΔLLR; Among them, LLR final is the final log-likelihood ratio, ΔLLR is the LLR residual correction term predicted by the neural network, and LLR MAP It means that the LLR value α output by the traditional Max-Log-MAP algorithm is the adaptive weight; α is calculated by the following formula: α=σ(W σ ·[SNR inst ,K]+b α ); σ is the Sigmoid function, which constrains α to be in the interval [0.3, 1.0]; W σ and b α is the weight matrix and bias term, which are learned through training.

7. A lightweight duobinary Turbo decoding method based on adaptive channel perception according to any one of claims 1 to 6, characterized in that: Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes; including: Design a dual-threshold early stopping strategy as follows: Among them, T is the total number of bits of a frame of data, δ is the confidence change threshold, softmax(LLR k ) indicates that the LLR value is normalized and converted into a probability distribution. represents the log-likelihood ratio of the t-th bit during the k-th iteration; represents the log-likelihood ratio of the t-th bit during the k-1-th iteration.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of the lightweight dual binary Turbo decoding method based on adaptive channel sensing according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the lightweight duo-binary Turbo decoding method based on adaptive channel sensing are implemented in any one of claims 1 to 7.

10. A lightweight duobinary turbo decoding system based on adaptive channel sensing, characterized in that: include: The data acquisition and preprocessing module is configured to: data acquisition and preprocessing; Data refers to the soft decision sequence after demodulation; Data preprocessing refers to converting the demodulated soft decision sequence into amplitude and phase information including symbol level; The duobinary turbo decoding module is configured to: input the preprocessed data into the trained decoding model to implement duobinary turbo decoding; The decoding model includes a dynamic channel perception module, a bidirectional GRU network, a multi-branch decoding network, a residual correction unit, and an iteration termination controller; including: In the dynamic channel perception module, a sliding window is used to extract the kurtosis, instantaneous signal-to-noise ratio, and zero-crossing rate features of the received signal; a bidirectional GRU network is used to classify the noise type; Dynamically activate the decoding branches in the multi-branch decoding network of corresponding complexity according to the noise type classification results, and fuse the traditional Max-Log-MAP algorithm output with the neural network prediction value through the residual correction unit; Dynamic optimization of the decoding process is achieved through an iterative termination controller based on confidence changes; The Adam optimizer is used to train the decoding model end-to-end, and the sigmoid function is used to make a hard decision on the final LLR value to output the decoding result.

Images