Polarization code optimization method and system suitable for dual selective fading channel

Abstract

The invention discloses a polarization code optimization method and system suitable for a dual-selective fading channel, and mainly solves the problems that an existing polarization code is poor in performance and complex in coding structure in the dual-selective fading channel. According to the scheme, the method comprises the steps that a Transform model is adopted as a reinforcement learning agent, and feature extraction and scoring are conducted on reliability distribution of all sub-channels; the intelligent agent generates a polarization code construction sequence based on the score vector; establishing a polarization code communication link as an interaction environment, inputting a polarization code construction sequence into the environment for interaction, and feeding back a frame error rate of current interaction to an intelligent agent as a performance index of the round; the intelligent agent calculates a reward signal according to the frame error rate and updates parameters of the intelligent agent in combination with a strategy gradient algorithm, so that the intelligent agent continuously optimizes a scoring strategy and a construction strategy of the intelligent agent; and repeating the steps until the output sequence of the intelligent agent converges step by step to obtain an optimal polarization code construction sequence. The polarization code construction complexity is low, the decoding performance is excellent, and the method can be used for signal transmission under a double selective fading channel.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and further relates to a polarization code optimization method and system, which can be used for signal transmission under a dual-selective fading channel. Background Art

[0002] Polar codes, proposed by Arikan in 2009, transform the polarization of a symmetric binary input channel and then decompose it into a set of subchannels with polarized reliabilities based on a subchannel sorting method. After polarizing the channel reliability, information bits are assigned to high-reliability subchannels and frozen bits are assigned to low-reliability subchannels to achieve asymptotic channel capacity. Commonly used subchannel sorting methods include Bhattacharyya parameter calculation based on density evolution (DE), Gaussian approximation (GA), and polarization weighting (PW). DE offers the highest accuracy, but the computational complexity of these methods increases exponentially with channel length, and each has its own drawbacks. For example, the GA method requires repeated parameter calculation for different signal-to-noise ratios when estimating actual channel reliability. The PW method, designed only for Gaussian channels, struggles to accurately estimate channel reliability in other complex channels.

[0003] Polar codes can approach the Shannon limit under ideal conditions. However, current traditional construction methods cannot accurately calculate the reliability of each subchannel in complex and changing channel environments, making it difficult to construct a coding sequence suitable for the current channel. To address this issue, the coding construction of polar codes can be viewed as a sequential decision problem. In recent years, reinforcement learning has demonstrated excellent dynamic learning capabilities for sequential decision problems. Furthermore, the Transformer model, with its global attention mechanism, can capture long-range dependencies and dynamic patterns from massive amounts of training data. Therefore, to further improve the transmission performance of polar codes in complex channels, a combination of reinforcement learning and the Transformer model can be used to estimate the reliability of each subchannel under real-time channel conditions. This approach can capture the state information of complex, time-varying channels, replacing expensive iterative density evolution or simulation sorting methods, enabling rapid online construction and optimizing the signal transmission performance of polar codes in doubly selective fading channels.

[0004] Patent document CN 201710660483.1 discloses a sequence generation and data decoding method and device. Based on different polar code lengths, number of information bits, code rate, and segmentation methods, it dynamically selects a bit index sequence with descending reliability from a pre-designed basic sequence library according to its nested characteristics. In subsequent encoding, the encoder uses this sequence as a guide to sequentially fill the information bits into the positions before polarization transformation and perform polarization encoding. However, this method has the following shortcomings: the predetermined sequence is not based on the real-time channel state, the calculation accuracy is low in extreme or non-Gaussian channel environments, and the optimal sequence requires offline design and simulation. If the channel model changes significantly, the sequence library needs to be regenerated.

[0005] Patent document CN 202010973768.2 discloses a method and apparatus for constructing an index sequence for polarization codes. The method comprises the following steps: obtaining an index set; obtaining multiple nested subsets of the index set; performing polarization coding on each of the multiple nested subsets to obtain multiple nested polarization codes; obtaining performance evaluation results for the multiple nested polarization codes; and, if the performance evaluation results for the multiple nested polarization codes do not meet a preset condition, updating the multiple nested subsets of the index set based on the performance evaluation results, repeating the above steps until the performance evaluation results meet the preset condition or the number of updates reaches a preset number; and obtaining an index sequence for the target polarization code based on the multiple nested subsets of the updated index set. Because this method employs nested subsets and cross-segmentation, and infers reliability ranking by generating and evaluating nested polarization codes of different sizes in parallel, it cannot correctly capture the long-range dependencies between all subchannels, making it difficult to accurately estimate the reliability of the subchannels. Summary of the Invention

[0006] The present invention aims to address the deficiencies of the prior art and propose a polar code optimization method suitable for dual-selective fading channels, so as to accurately estimate the reliability of polarization sub-channels under dual-selective fading channels.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] 1. A polar code optimization method for dual-selective fading channels, comprising:

[0009] (1) The Transformer model is used as a reinforcement learning agent, and its multi-head self-attention mechanism is used to perform deep feature extraction and scoring on the reliability distribution of each sub-channel to which the polar code belongs, thereby capturing the interdependence between different sub-channels.

[0010] (2) Based on the scoring results of the Transformer model, the agent uses the exploration strategy to adjust the position of the candidate sequence and generate a new coding construction sequence;

[0011] (3) Build a full communication transmission and reception link including polarimetric encoding and decoding, and use it as an environment for interaction with the intelligent agent. Then, input the coding construction sequence into the environment for encoding, channel simulation, and decoding interaction. Then, the frame error rate generated by the current interaction is fed back to the intelligent agent as a performance indicator for this round of interaction.

[0012] (4) The agent constructs a reward function based on the frame error rate feedback from the interactive environment, uses this reward function to map a lower frame error rate to a higher positive reward, and combines it with the policy gradient algorithm to update the parameters of the Transformer model, so that it can continuously optimize its own scoring strategy;

[0013] (5) The updated Transformer model is used as the strategy network for the next round of interaction to re-extract features and score the reliability distribution of each sub-channel;

[0014] (6) Repeat the above (2) to (5) until the output sequence of the agent gradually converges and the optimal coding construction sequence is obtained.

[0015] Furthermore, the multi-head self-attention mechanism of the Transformer model in (1) is used to perform deep feature extraction and scoring on the reliability distribution of each sub-channel to which the polar code belongs, so as to capture the mutual dependence between different sub-channels. The implementation thereof includes:

[0016] (1a) Mapping the sub-channel reliability vectors of the polar code to the matrix representations of query (Q), key (K), and value (V);

[0017] (1b) Set up multiple attention heads in parallel, calculate the scaled dot product of all attention heads, and obtain the attention weight matrix through the Softmax layer;

[0018] (1c) Concatenate the weighted sums of the attention heads and fuse them into a unified feature representation through linear mapping;

[0019] (1d) Perform residual connection and layer normalization on the fused features to enhance the stability of the model during training;

[0020] (1e) Use a feedforward fully connected network for nonlinear transformation, and again use residual connections and layer normalization to generate the final sub-channel reliability score vector;

[0021] (1f) Based on the reliability score vector of the sub-channel, the similarity between each sub-channel is calculated and compared to obtain the mutual dependence relationship between different sub-channels.

[0022] Furthermore, the agent in (4) constructs a reward function based on the frame error rate fed back by the interactive environment, and uses the reward function to map a lower frame error rate to a higher positive reward, which is implemented as follows:

[0023] (4a) Let the frame error rate FER obtained by the agent in the tth round of interaction be t , define its basic reward R t ;

[0024] (4b) According to the preset reference frame error rate FER ref , define its reference reward R ref ;

[0025] (4c) Calculate the basic reward R t With reference reward R ref The difference ΔR t :

[0026] (4d) According to the reward difference ΔR t , using a linear scaling strategy to construct the initial reward function r t (0) ,

[0027] (4e) Apply a nonlinear transformation to the initial reward function to amplify the reward difference r corresponding to different frame error rates t (1) ;

[0028] (4f) The reward function r after nonlinear transformation t (1) Perform smoothing to obtain the smoothed reward function r t (sm)

[0029] Furthermore, the parameters of the Transformer model are updated in combination with the policy gradient algorithm in (4), and its implementation includes:

[0030] (4g) Within a training batch, the agent interacts with the environment multiple times, and records the construction sequence generated by the Transformer model and the corresponding reward in each interaction;

[0031] (4h) Based on the policy gradient algorithm, the product of the probability of selection of each constructed sequence under the current strategy and the reward is calculated as the gradient estimator;

[0032] (4i) Take the average of the gradient estimators of all sequences within a training batch to obtain the overall policy gradient direction;

[0033] (4j) Based on the overall policy gradient direction, the adaptive optimizer is used to update the parameters of the Transformer model.

[0034] 2. A polar code optimization system for dual-selective fading channels, comprising:

[0035] The scoring module is used to use the multi-head self-attention mechanism of the Transformer model to perform deep feature extraction on the initial reliability indicators of each sub-channel obtained by simulation and output the sub-channel score vector;

[0036] A sequence generation module is used to normalize the score vector and map it to position selection probabilities. It then uses an exploration strategy to construct an initial position sequence. This initial position sequence is then fine-tuned according to predefined swapping, insertion, or rearrangement rules to obtain a polar code construction sequence.

[0037] The environment interface module is used to pass the polar code construction sequence into the dynamic link library for simulation and output the frame error rate of the current polar code;

[0038] A reward function calculation module is used to compare the frame error rate output by the environment interface module with a preset reference frame error rate, and obtain a reward function after performing nonlinear transformation and smoothing operations on the calculation result;

[0039] The parameter update module is used to accumulate multiple interaction data and calculate the gradient estimator according to the reward function and policy gradient algorithm in a single batch, and then update the model parameters with the help of the adaptive optimizer.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] First, the present invention utilizes a Transformer encoder model to perform multi-dimensional scoring of the reliability distribution of each subchannel. By fusing context vectors from different attention heads, it accurately captures the global dependencies between subchannels. Compared to existing local evaluations based solely on a single weight or mutual information metric, the present invention exploits the hidden global correlations between polarimetric subchannels. This allows for accurate estimation of the reliability of polarimetric subchannels in dual-selective fading channels, generating more refined and adaptable polar code construction sequences.

[0042] Secondly, the present invention introduces real decoding performance into closed-loop feedback, and realizes online updating of Transformer model parameters through the policy gradient algorithm and adaptive optimizer in reinforcement learning. This allows the system to automatically adjust the sorting strategy of the sub-channel sequence after each interaction between the intelligent agent and the environment without relying on any preset static rules, and respond in real time to dynamic changes such as multipath and deep fading in the doubly selective fading channel, significantly improving the matching degree and robustness of the polar code construction sequence and decoding performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is an implementation scenario diagram of the present invention;

[0044] Figure 2 This is a general flow chart of the polar code optimization method applicable to dual-selective fading channels of the present invention;

[0045] Figure 3 This is a sub-flowchart of polar code reliability distribution depth feature extraction and scoring in the method of the present invention;

[0046] Figure 4 Schematic diagram of the full communication transmission and reception link of polar coding in the method of the present invention;

[0047] Figure 5 This is a block diagram of a polar code optimization system applicable to dual-selective fading channels according to the present invention;

[0048] Figure 6 This is a simulation chart comparing the polar code performance of the present invention and the prior art under a message length of 32 and a code length of 64.

[0049] Figure 7 This is a simulation diagram comparing the polar code performance of the present invention and the prior art under an information length of 96 and a code length of 192. DETAILED DESCRIPTION

[0050] The embodiments and effects of the present invention are further described in detail below with reference to the accompanying drawings.

[0051] Reference Figure 1 The present invention is implemented based on the intelligent agent training scenario and the interactive environment scenario.

[0052] The agent training scenario uses subchannel prior reliability information and the reward signal from the previous round of interaction as input. The agent uses the multi-head self-attention mechanism of the Transformer model and the policy network operation to generate the current polar code construction sequence and output the construction sequence to the interaction environment scenario.

[0053] The interactive environment scenario uses the polar code construction sequence as input, sequentially performs polar coding, Walsh spreading, dual-selective fading channel simulation, Rake reception, and polar decoding to obtain a frame error rate performance indicator, and feeds the performance indicator back to the intelligent agent scenario.

[0054] Based on the above scenario, implementation examples of the present invention include:

[0055] Example 1: Polar Code Optimization Method for Dual Selective Fading Channels

[0056] Reference Figure 2 The implementation steps of this example include the following:

[0057] Step 1: Initialize the channel and Transformer model parameters.

[0058] The Transformer model is an existing AI model whose function is to convert the original feature vector into a high-dimensional representation through linear mapping, and superimpose position encoding on it to maintain the relative order of elements, and then use multi-head self-attention in parallel to capture global dependencies and project them to the target dimension, and finally fine-tune the network weights based on gradient descent or adaptive optimizer.

[0059] In this example, the dual selective fading channel is initialized and its parameters are set, including the number of multipath components P, the path delay τ p , channel fading coefficient h p and Gaussian white noise power N0;

[0060] Initialize Transformer model parameters, including learning rate lr and embedding dimension d model , number of attention heads H, hidden dimension d of feedforward network ff , number of layers L.

[0061] Step 2: Map the polarization subchannel reliability vector into a high-dimensional vector.

[0062] 2.1) Using the existing construction method under Gaussian channel, a polarization channel reliability vector of length N is simulated: r = [r1, r2, ..r i ..,r N ] Τ , and perform min-max normalization operation on it to obtain the normalized vector

[0063]

[0064] in, r i represents the reliability of the i-th polarization channel, r min 、r max are the minimum and maximum values ​​of the collected reliability respectively. It can eliminate dimensional differences and improve network training stability:

[0065] 2.2) The normalized vector Mapped to d by linear transformation model dimensional feature space X:

[0066]

[0067] Among them, the dimension of the feature space X is It is a trainable parameter. ReLU is a neural network activation function.

[0068] 2.3) Based on the existing "sine-cosine" position encoding scheme, construct the position encoding matrix PE:

[0069] PE=[PE1,..,PE i ,..,PE N ],

[0070] Where i = 1, 2, ..., N, PE i is the position encoding vector of the i-th subchannel, PE i The dimension is d model , remember PE i The sequence number of the paired elements in is k, and the elements in the kth pair are calculated:

[0071]

[0072] in, PE i [2k] is vector PE i The 2kth element in PE i [2k+1] is the vector PE i The 2k+1th element in

[0073] 2.4) The feature space X is superimposed element by element with the position encoding matrix PE to obtain the attention feature matrix X′ of the polarized sub-channel:

[0074] X′=X+PE.

[0075] Step 3: Generate a scoring vector for the polarization subchannel reliability.

[0076] The multi-head self-attention mechanism in the Transformer model is used to extract the features of the polarized sub-channels and perform weighted fusion. The fusion results are then nonlinearly mapped through a feedforward network to output the sub-channel reliability score vector.

[0077] Reference Figure 3 , this step is implemented as follows:

[0078] 3.1) Extract the features of the sub-channel reliability vector:

[0079] 3.1.1) Map the feature matrix X′ after superimposing position encoding into a matrix representation of query (Q), key (K) and value (V):

[0080] Q=X′W Q , K=X′W K , V=X′W V ,

[0081] in, is a trainable projection weight matrix;

[0082] 3.1.2) Let the number of self-attention heads in the Transformer model be H. Each self-attention head performs a scaled dot product operation on the query (Q), key (K) and value (V) and outputs an attention weight matrix. The attention weight matrix output by the hth self-attention head is denoted as head h :

[0083]

[0084] Where h = 1, 2, ..., H, d k For a single head dimension, softmax is a normalized exponential function that normalizes the input vector and maps it to a probability distribution.

[0085] 3.2) Project the score vector of the output polarimetric sub-channel:

[0086] 3.2.1) Concatenate the weight matrix head output by each attention head h , get the extended feature matrix H concat :

[0087] H concat =[head1,..,head h ,..,head H ];

[0088] 3.2.2) For the extended feature matrix H concat Perform linear mapping and add offset values ​​to achieve dimensionality reduction fusion of multi-dimensional features and obtain the fusion output matrix U:

[0089] U=H concat W O +b O ,

[0090] Among them, W O is the linear mapping weight matrix, b O is the offset vector;

[0091] 3.2.3) Perform residual connection on the fusion output matrix U and the feature matrix X′ to obtain the connected matrix R:

[0092] R=U+X′;

[0093] 3.2.4) Use the existing LayerNorm function to normalize the matrix R and obtain the normalized result Z:

[0094] Z = LayerNorm(R);

[0095] 3.2.5) Input the normalized result Z into a two-layer feedforward fully connected network and activate it using the ReLU activation function to obtain the intermediate feature matrix Y before residual connection and normalization:

[0096] Y=ReLU(ZW1+b1)W2+b2,

[0097] Among them, W1 and W2 are the linear mapping weight matrices of the first and second layers respectively, and b1 and b2 are the offset vectors of the first and second layers respectively;

[0098] 3.2.6) Perform a residual connection on the intermediate feature matrix Y and normalize it using the existing LayerNorm function to generate the reliability score vector S of the polarimetric subchannel:

[0099] S=[s1,..,s i ,..,s N ]=LayerNorm(Y+Z),

[0100] Among them, s i is the score value of the i-th subchannel, i=1,2,...,N.

[0101] Step 4: Generate a polar code construction sequence based on the scoring vector.

[0102] 4.1) Use the Softmax layer to evaluate the reliability score vector S = [s1, .., s i ,..,s N ] is normalized to obtain a set of sub-channel position selection probability distribution P:

[0103] P=[p1,..,p i ,..,p N ],

[0104] Among them, p i is the selection probability of the i-th subchannel position, which is calculated as follows:

[0105]

[0106] 4.2) Based on the preset exploration rate ε, the sampling operation is performed on the positions of all sub-channels according to the selection probability distribution P to obtain a set of candidate positions

[0107] 4.3) For collections For each candidate position i, according to its selection probability p i And the existing exploration strategy, using predefined fine-tuning operations, including position replacement, element insertion, index relocation, etc., to The order of candidate positions in the CNN is locally adjusted;

[0108] 4.4) The candidate positions after adjustment are used as a new polar code construction sequence.

[0109] Step 5: Pass the polarization code construction sequence into the agent interaction environment.

[0110] A full communication link, including polarimetric encoding and decoding, is built and used as an environment for interaction with the agent. The encoded sequence is then input into this environment for encoding, channel simulation, and decoding. Finally, the frame error rate generated by the current interaction is fed back to the agent as a performance indicator for this round of interaction.

[0111] Reference Figure 4 , this step is implemented as follows:

[0112] 5.1) Implement the full communication link including polarization coding, Walsh spreading, digital modulation, channel emulation, Rake reception, and soft decision decoding using C language, following the uplink and downlink physical channel processing procedures defined in 3GPP TS 38.212.

[0113] 5.2) At the transmitter, a sequence is constructed based on the polar code output by the Transformer model, polarization encoding is performed on the original information, and a polar code codeword of length N is generated;

[0114] 5.3) Multiply each symbol in the polar code word by a preset Walsh matrix to obtain a spread spectrum sequence;

[0115] 5.4) Perform BPSK or higher-order mapping on each spread symbol in the spread sequence to map the discrete symbol into the corresponding complex baseband signal;

[0116] 5.5) The complex baseband signal is transmitted into a doubly selective fading channel, where channel convolution and noise addition are performed to generate multipath signal components.

[0117] 5.6) At the receiving end, a Rake receiver is used to perform weighted combination of the multipath signal components to obtain the log-likelihood ratio vector of the received signal, which is then input into a polarimetric decoder for decoding.

[0118] 5.7) Compare the decoded output with the original information bit by bit and calculate the frame error rate, which is then fed back to the agent.

[0119] Step 6: The agent calculates the reward function based on the frame error rate.

[0120] 6.1) Assume that the frame error rate (FER) obtained by the agent in the tth round of interaction is t , based on the frame error rate, the basic reward is defined as r t :

[0121] rt =-log2(FER t );

[0122] 6.2) According to the preset reference frame error rate FER ref , define its reference reward as r ref :

[0123] r ref =-log2(FER ref );

[0124] 6.3) Calculate the basic reward r t With reference reward r ref The difference Δr t :

[0125] Δr t =r t -r ref ;

[0126] 6.4) According to the reward difference Δr t , using a linear scaling strategy to construct the initial reward function

[0127]

[0128] Where α is the scaling factor, when FER t <FER ref When r t >r ref , the initial reward function

[0129] 6.5) For the initial reward function Apply nonlinear transformation to amplify the reward differences corresponding to different frame error rates:

[0130]

[0131] Among them, the exponent γ>1 is adjusted according to the needs to enhance the resolution capability of the low frame error rate range. is the reward function after nonlinear transformation;

[0132] 6.6) According to the reward function of the previous round And the preset smoothing coefficient λ, for the current reward function Perform smoothing to obtain the smoothed reward function

[0133]

[0134] Among them, λ∈[0,1], is the initial reward function.

[0135] Step 7: Update the parameters of the Transformer model.

[0136] 7.1) In a single training batch, assume that the agent interacts with the exchange environment M times. For the i-th interaction, record the polar code construction sequence a output by the Transformer model. (i) And the corresponding reward signal R (i) , and stored in the interactive trajectory buffer;

[0137] 7.2) Based on the policy gradient principle, calculate the policy gradient estimator g for each record in the interaction trajectory buffer (i) :

[0138]

[0139] Among them, π θ (a (i) ) is the Transformer model that generates sequence a under the current parameter θ (i) The probability of selection;

[0140] 7.3) For all M gradient estimators Perform arithmetic averaging to obtain the overall policy gradient direction of the training batch

[0141]

[0142] 7.4) Call the adaptive optimizer according to the overall strategy gradient direction Adjust and update the Transformer model parameters θ:

[0143]

[0144] Among them, α is the learning rate and θ′ is the updated parameter.

[0145] Step 8: Repeat steps 3 to 7 above. When the reward signal R (i) The fluctuation amplitude in multiple consecutive rounds of interaction does not exceed the preset threshold ε grad , or the number of model training times reaches the preset maximum value T max When , the model training process is terminated and the final optimized polar code construction sequence is output.

[0146] It should be noted that the flowchart representations or method representations of the above embodiments can be understood as representing a module, segment, or portion of code that includes one or a group of executable instructions configured to implement a specific logical function or process step. The present invention is not limited to the disclosed preferred embodiments, and its implementation may not follow the order presented or discussed. That is, the step numbers in the present specification and claims are merely for the purpose of clarifying the embodiments of the present invention and facilitating understanding, and the order of the numbering is not limiting.

[0147] Example 2: Polar Code Optimization System for Dual Selective Fading Channels

[0148] Reference Figure 5 The polar code optimization system in this example includes: a scoring module 1, a construction sequence generation module 2, an environment interface module 3, a reward function calculation module 4, and a parameter update module 5. The scoring module 1 includes: a preprocessing submodule 11, an embedding mapping submodule 12, a multi-head self-attention submodule 13, a feature fusion submodule 14, and an output mapping submodule 15. The working principle of the entire system is as follows:

[0149] The scoring module 1 is used to perform deep feature extraction on the initial reliability indicators of each sub-channel obtained by simulation using the multi-head self-attention mechanism of the Transformer model, wherein: the pre-processing sub-module 11 is used to receive the initial reliability indicators of each sub-channel obtained by simulation, and perform normalization processing to eliminate dimensional differences and improve data stability, and pass the pre-processed reliability indicators to the embedding mapping sub-module 12, the embedding mapping sub-module 12 is used to map the pre-processed reliability indicators to a high-dimensional vector space through a linear transformation to form three input representations of query (Q), key (K), and value (V), and pass them to the multi-head self-attention mechanism. The head self-attention submodule 13 is used to parallelly calculate the scaled dot product and Softmax weight in each attention head to obtain a weighted context vector that characterizes the mutual dependence between sub-channels, and pass it to the feature fusion submodule 14. The feature fusion submodule 14 is used to fuse the context vectors of each attention head and stabilize its features through residual connection and layer normalization. The context vectors are then passed to the output mapping submodule 15. The output mapping submodule 15 is used to further refine the fused high-dimensional features using a feedforward fully connected network, project them into the final sub-channel score vector, and pass it to the construction sequence generation module 2;

[0150] The construction sequence generation module 2 is used to normalize the score vector and map it into position selection probabilities, and construct an initial position sequence based on the exploration strategy. The initial position sequence is fine-tuned according to predefined exchange, insertion, or rearrangement rules to obtain a polar code construction sequence, and the polar code construction sequence is transmitted to the environment interface module 3.

[0151] The environment interface module 3 is used to input the polar code construction sequence into the dynamic link library for simulation, output the frame error rate of the current polar code, and transmit it to the reward function calculation module 4;

[0152] The reward function calculation module 4 is used to compare and calculate the frame error rate output by the environment interface module with a preset reference frame error rate, obtain a reward function after performing nonlinear transformation and smoothing operations on the calculation result, and transmit the reward function to the parameter update module 5;

[0153] The parameter updating module 5 is used to accumulate multiple interaction data and calculate gradient estimates according to the reward function and policy gradient algorithm in a single batch, and then update the model parameters with the help of the adaptive optimizer.

[0154] It should be noted that the above-mentioned functional modules can be implemented in whole or in part through software, hardware, firmware, or any combination thereof. When implemented using software, they can be implemented in whole or in part in the form of a program instruction product. The program instruction product includes one or a group of program instructions. When the program instructions are loaded and executed on a computer, the process or function described is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium, or transferred from a computer-readable and writable storage medium to another computer-readable and writable storage medium.

[0155] The direct coupling or communication connection between the modules shown or discussed in this embodiment can be achieved through indirect coupling or communication connection of some interfaces, devices or modules. The various functional modules and submodules in this embodiment can be dynamically located in a processing component, or each module can exist physically separately, or two or more modules can be dynamically located in a processing component. When the above-mentioned dynamic components are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. The storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0156] The effect of the present invention can be further illustrated by the following simulation experiments:

[0157] 1. Simulation Experiment Conditions

[0158] The simulation experiment in this embodiment is completed on the Windows 11 64-bit platform. The hardware environment includes Intel's 12th-generation Core i7 processor and NVIDIA GeForce RTX 3060 GPU. The software environment uses Python 3.9 to implement the Transformer reinforcement learning agent on the PyCharm IDE, and uses C language to implement the full communication transmission and reception link including polarization encoding and decoding on Visual Studio 2022, and the two are collaboratively simulated through Python's ctypes interface.

[0159] The key parameters of the Transformer reinforcement learning agent are: layer =4, hidden dimension d model =128, number of attention heads H = 8, learning rate α = 1×10 -4 , exploration rate ε = 0.1, feedforward network intermediate dimension d ff =512, reward smoothing coefficient α smooth =0.9, policy gradient batch size M=64.

[0160] 2. Simulation Content

[0161] Simulation 1: Under the above simulation experimental conditions, the polar code information length K is set to 32, and the polar code codeword length E is set to 64. The polar code construction sequence of the existing 3GPP TS 38.212 standard is used as the reference sequence, and the polar code optimization method described in this invention is used to generate the optimized polar code construction sequence. Under the same signal-to-noise ratio conditions, the reference sequence and the optimized sequence are sequentially input into the communication transceiver simulation environment. The end-to-end encoding-channel simulation-decoding-performance evaluation process is executed, and the frame error rate (FER) of each sequence is calculated. The results are as follows: Figure 6 shown.

[0162] Simulation 2: Under the above simulation experimental conditions, the polar code information length K is set to 96, and the polar code codeword length E is set to 192. The polar code construction sequence of the existing 3GPP TS 38.212 standard is used as the reference sequence, and the polar code optimization method described in the present invention is used to generate the optimized polar construction sequence. The end-to-end simulation experiment is repeated under the same steps and conditions, and the frame error rate (FER) of each is calculated. The results are as follows: Figure 7 shown.

[0163] from Figure 6 、 Figure 7 As can be seen, the optimized polar codes of the present invention exhibit lower frame error rates at all tested signal-to-noise ratio points compared to the polar codes of the existing 3GPP TS 38.212 standard. This verifies that the polar code optimization method has superior decoding performance and good practicality in a dual-selective fading channel environment.

[0164] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, it is possible to make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A polar code optimization method for dual-selective fading channels, characterized in that: include: (1) The Transformer model is used as a reinforcement learning agent, and its multi-head self-attention mechanism is used to perform deep feature extraction and scoring on the reliability distribution of each sub-channel to which the polar code belongs, thereby capturing the interdependence between different sub-channels. (2) Based on the scoring results of the Transformer model, the agent uses the exploration strategy to adjust the position of the candidate sequence and generate a new coding construction sequence; (3) Build a full communication transmission and reception link including polarimetric encoding and decoding, and use it as an environment for interaction with the intelligent agent. Then, input the coding construction sequence into the environment for encoding, channel simulation, and decoding interaction. Then, the frame error rate generated by the current interaction is fed back to the intelligent agent as a performance indicator for this round of interaction. (4) The agent constructs a reward function based on the frame error rate feedback from the interactive environment, uses this reward function to map a lower frame error rate to a higher positive reward, and combines it with the policy gradient algorithm to update the parameters of the Transformer model, so that it can continuously optimize its own scoring strategy; (5) The updated Transformer model is used as the strategy network for the next round of interaction to re-extract features and score the reliability distribution of each sub-channel; (6) Repeat the above (2) to (5) until the output sequence of the agent gradually converges and the optimal coding construction sequence is obtained.

2. The method according to claim 1, characterized in that The Transformer model in (1) is an existing AI model, and its functions include: Map the original feature vector to a high-dimensional feature vector; Add position information to high-dimensional feature vectors to maintain the relative order of elements in the vector; It uses its own multi-head self-attention mechanism to capture the global dependencies of high-dimensional vectors in parallel and project them into scores or feature vectors of the target dimension; Fine-tune its own network weights through gradient descent algorithm or adaptive optimizer.

3. The method according to claim 1, characterized in that In (1), the multi-head self-attention mechanism of the Transformer model is used to perform deep feature extraction and scoring on the reliability distribution of each sub-channel to which the polar code belongs, so as to capture the mutual dependence between different sub-channels. The implementation thereof includes: (1a) Mapping the sub-channel reliability vectors of the polar code to the matrix representations of query (Q), key (K), and value (V); (1b) Set up multiple attention heads in parallel, calculate the scaled dot product of all attention heads, and obtain the attention weight matrix through the Softmax layer; (1c) Concatenate the weighted sums of the attention heads and fuse them into a unified feature representation through linear mapping; (1d) Perform residual connection and layer normalization on the fused features to enhance the stability of the model during training; (1e) Use a feedforward fully connected network for nonlinear transformation, and again use residual connections and layer normalization to generate the final sub-channel reliability score vector; (1f) Based on the reliability score vector of the sub-channel, the similarity between each sub-channel is calculated and compared to obtain the mutual dependence relationship between different sub-channels.

4. The method according to claim 1, wherein In (2), the agent adjusts the position of the candidate sequence based on the scoring results of the Transformer model using an exploration strategy to generate a new coding construction sequence, which is implemented as follows: (2a) Normalize the sub-channel score vector output by the Transformer model to obtain the probability distribution of each position being selected; (2b) Randomly sample each subchannel position according to the preset exploration rate ε to obtain a set of candidate positions; (2c) For each candidate position, perform adjustment operations based on its probability distribution size and exploration strategy, including replacing, inserting or relocating the subchannel to the specified position index; (2d) The candidate positions after adjustment are used as the new coding construction sequence.

5. The method according to claim 1, wherein The construction of a full communication transmission and reception link including polarization coding in (3) includes: (3a) Develop the full communication link including polar coding and transmission in C language according to the existing 3GPP TS 38.212 communication protocol; (3b) Polar coding the original information according to the coding construction sequence at the transmitting end; (3c) performing Walsh spreading on the polarization-encoded codeword, spreading the symbols in the codeword according to a preset Walsh matrix to obtain a spreading sequence; (3d) digitally modulating each symbol in the spread spectrum sequence to convert it into a digital signal for channel transmission; (3e) Build a dual-selective fading channel model and add Gaussian white noise, then input the digital signal into the channel model to simulate the transmission process of the digital signal in the dual-selective fading channel; (3f) At the receiving end, a Rake receiver is used to capture and combine the multipath signal components in the dual-selective fading channel, output the log-likelihood ratio information of the signal, and use the log-likelihood ratio information to perform soft decision decoding to obtain a polarization decoding result; (3g) Calculating the difference between the polar decoding result and the original information to evaluate the performance of the current polar code; At this point, a complete communication transmission and reception link covering encoding, channel simulation, decoding and performance evaluation is formed.

6. The method according to claim 1, characterized in that In (3), the coding sequence is input into the agent interaction environment for coding, channel simulation, and decoding interaction, which is implemented as follows: (3h) Encapsulate the entire communication transmission and reception link, including polarization coding, channel simulation, and polarization decoding functions, into a dynamic link library, and reserve a unified interface function for external use; (3i) Convert the coded structure sequence output by the agent into an array format compatible with the C language; (3j) By calling the interface function, the converted array is passed to the dynamic link library, and then encoding, channel simulation and decoding are performed on the converted array in sequence to complete the interaction between the intelligent agent and the environment.

7. The method according to claim 1, characterized in that In (4), the agent constructs a reward function based on the frame error rate fed back by the interactive environment, and uses the reward function to map a lower frame error rate to a higher positive reward. The implementation includes: (4a) Let the frame error rate FER obtained by the agent in the tth round of interaction be t , define its basic reward as: R t =-log2(FER t ); (4b) According to the preset reference frame error rate FER ref , define its reference reward as: R ref =-log2(FER ref ); (4c) Calculate the basic reward R t With reference reward R ref The difference ΔR t : ΔR t =R t -R ref ; (4d) According to the reward difference ΔR t , using a linear scaling strategy to construct the initial reward function r t (0) : r t (0) =αΔR t =α(R t -R ref ), Where α is the scaling factor, when FER t <FER ref When R t >R ref , the initial reward function r t (0) >0; (4e) Apply a nonlinear transformation to the initial reward function to amplify the reward differences corresponding to different frame error rates: r t (1) =sign(r t (0) )|r t (0) | γ , The exponent γ is greater than 1 and is adjusted according to the demand to enhance the ability to distinguish low frame error rate intervals. (4f) The reward function r after nonlinear transformation t (1) To perform smoothing: where λ∈[0,1] is the smoothing coefficient and 8. The method according to claim 1, characterized in that In (4), the parameters of the Transformer model are updated by combining the policy gradient algorithm, and its implementation includes: (4g) Within a training batch, the agent interacts with the environment multiple times, and records the construction sequence generated by the Transformer model and the corresponding reward in each interaction; (4h) Based on the policy gradient algorithm, the product of the selection probability and reward of each constructed sequence under the current strategy is calculated as the gradient estimator; (4i) Average the gradient estimators of all sequences within a training batch to obtain the overall policy gradient direction; (4j) Based on the overall policy gradient direction, the adaptive optimizer is used to update the parameters of the Transformer model.

9. A polar code optimization system for dual-selective fading channels, characterized in that: include: The scoring module is used to use the multi-head self-attention mechanism of the Transformer model to perform deep feature extraction on the initial reliability indicators of each sub-channel obtained by simulation and output the sub-channel score vector; A sequence generation module is used to normalize the score vector and map it to position selection probabilities. It then uses an exploration strategy to construct an initial position sequence. This initial position sequence is then fine-tuned according to predefined swapping, insertion, or rearrangement rules to obtain a polar code construction sequence. The environment interface module is used to pass the polar code construction sequence into the dynamic link library for simulation and output the frame error rate of the current polar code; A reward function calculation module is used to compare the frame error rate output by the environment interface module with a preset reference frame error rate, and obtain a reward function after performing nonlinear transformation and smoothing operations on the calculation result; The parameter update module is used to accumulate multiple interaction data and calculate the gradient estimator according to the reward function and policy gradient algorithm in a single batch, and then update the model parameters with the help of the adaptive optimizer.

10. The system according to claim 9, characterized in that The scoring module includes: The pre-processing sub-module is used to receive the initial reliability indicators of each sub-channel obtained by simulation and perform normalization processing to eliminate dimensional differences and improve data stability; The embedding mapping submodule is used to map the pre-processed reliability indicators into a high-dimensional vector space through linear transformation to form three input representations: query (Q), key (K), and value (V); The multi-head self-attention submodule is used to parallelize the scaled dot products and softmax weights in each attention head to obtain a weighted context vector that represents the interdependence between sub-channels. The feature fusion submodule is used to concatenate and fuse the context vectors of each attention head and stabilize their features through residual connections and layer normalization; The output mapping submodule is used to further refine the fused high-dimensional features using a feedforward fully connected network and project them into the final sub-channel score vector.

Images