BCH code parallel decoding method based on artificial intelligence

Abstract

The invention relates to the field of wireless communication networks, and provides a BCH (Broadcast Channel) code parallel decoding method based on artificial intelligence, which comprises the following steps of: acquiring a to-be-decoded BCH receiving code word from a physical channel; inputting a to-be-decoded BCH receiving code word into the trained error pattern prediction model, outputting a predicted error pattern, and performing error correction on the to-be-decoded BCH receiving code word by using the predicted error pattern to obtain an estimated sending code word; respectively carrying out syndrome calculation on the BCH receiving code word to be decoded and the estimated sending code word so as to judge the validity of the code word; and based on a result of syndrome calculation, carrying out dual-channel decoding control according to a sequence of firstly judging a BCH receiving code word to be decoded and then judging an estimated sending code word, carrying out parallel prediction on an error pattern of the BCH receiving code word by introducing an artificial intelligence model, bypassing a serial iterative calculation bottleneck in a traditional BCH decoding process, and carrying out decoding on the error pattern of the BCH receiving code word. Error position determination can be completed in a single forward reasoning process, decoding time delay is effectively reduced, and throughput is improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication networks, and in particular to a parallel decoding method for BCH codes based on artificial intelligence. Background Technology

[0002] BCH (Bose-Chaudhuri-Hocquenghem) codes are a classic type of cyclic error-correcting code. Due to their strong error-correcting capabilities and clear structure, they are widely used in scenarios with extremely high reliability requirements, such as 5G communication, satellite communication, storage devices, deep space communication, and spacecraft control systems. These scenarios typically face harsh channel conditions and high data throughput requirements, making BCH codes an important choice due to their excellent error-correcting performance.

[0003] Traditional BCH decoding algorithms, especially the BM (Berlekamp-Massey) algorithm for calculating the error location polynomial and the Chien Search algorithm for locating the error location, are inherently serial and iterative. Each step depends on the result of the previous step, leading to inefficient decoding. This strong serial dependency, requiring each step to be executed sequentially, severely limits decoding speed, making it difficult to meet the low latency and ultra-high throughput requirements of high-throughput satellites.

[0004] To address this bottleneck, existing technologies attempt to accelerate the decoding process through hardware parallelization (such as parallel Chan's search) or algorithm optimization (such as the improved BM algorithm). However, these optimization schemes have failed to fundamentally break through the serial iterative computation paradigm, resulting in performance improvement bottlenecks. Therefore, there is an urgent need in this field for a novel decoding architecture capable of resolving the performance bottleneck of traditional BCH decoding. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems by proposing an artificial intelligence-based parallel decoding method for BCH codes. This method introduces a deep neural network, aiming to learn a complex mapping from the received codeword to the most likely error pattern. The AI ​​model can directly predict the error pattern, fundamentally bypassing the serial iterative steps in the traditional decoding process, thus achieving parallelization and ultra-low latency in the decoding process. Simultaneously, a dual-channel architecture of AI path and traditional verification path is constructed, ensuring system reliability on one hand and endowing the system with online evolution capabilities on the other. The objective of this invention can be achieved through the following technical solutions: This invention provides a parallel decoding method for BCH codes based on artificial intelligence, including: Step S1: Obtain the BCH receive codeword to be decoded from the physical channel; Step S2: Input the BCH received codeword to be decoded into the trained error pattern prediction model, output the predicted error pattern, and use the predicted error pattern to correct the error of the BCH received codeword to be decoded to obtain the estimated transmitted codeword. Step S3: Perform synod calculations on the BCH received codeword to be decoded and the estimated transmitted codeword to determine the validity of the codeword; based on the result of the synod calculation, perform dual-channel decoding control in the order of first determining the BCH received codeword to be decoded and then determining the estimated transmitted codeword, and start the traditional BCH decoding process to supervise and update the model when the prediction fails.

[0006] Further, step S1 includes: The receiving antenna captures the modulated high-frequency analog signal from the physical channel, and performs down-conversion and filtering on the high-frequency analog signal to obtain the corresponding intermediate frequency analog signal; The intermediate frequency analog signal is converted from analog to digital to obtain a discrete time series digital signal. After carrier synchronization and symbol synchronization processing, the digital signal is mapped onto the corresponding constellation diagram to obtain the corresponding binary code stream. Based on the frame synchronization information, a complete BCH codeword length data segment r is extracted from the binary bitstream as the BCH receive codeword to be decoded.

[0007] Furthermore, in step S2, the error pattern prediction model is a deep neural network architecture that has been pre-built and trained. The deep neural network architecture is based on the mapping relationship between the BCH received codewords and the corresponding error patterns learned in the pre-training stage.

[0008] Furthermore, through a pre-built and trained deep neural network architecture, the deep neural network architecture is based on the mapping relationship between the BCH received codewords and the corresponding error patterns learned in the pre-training stage, including: A deep neural network architecture for constructing an error pattern prediction model is proposed. The deep neural network structure consists of an input layer, at least three one-dimensional convolutional layers, at least two one-dimensional max-pooling layers, a flattening layer, a fully connected layer, and an output layer connected sequentially. The input layer receives data of dimension 1. A one-dimensional sequence, Let be the length of the BCH codeword, and 1 represent one input feature corresponding to each bit position. A one-dimensional sequence is input to a first-dimensional convolutional layer. After convolution processing, the output feature of the first-dimensional convolutional layer is passed to a first-dimensional max-pooling layer for downsampling. The output of the first-dimensional max-pooling layer serves as the input to a second-dimensional convolutional layer. The second-dimensional convolutional layer convolutions the pooled features and passes the output feature to the second-dimensional max-pooling layer for further downsampling. The output of the second-dimensional max-pooling layer is passed to a third-dimensional convolutional layer. After extracting high-level features, the convolution result is input to a flattening layer for dimensionality expansion. The flattened features are then input to a fully connected layer for feature fusion and comprehensive decision-making. The output of the fully connected layer is finally passed to the output layer, where an activation function generates a bit-level error probability corresponding to the BCH codeword length, indicating the probability of an error occurring at each bit position. Construct a pre-trained dataset and pre-train the error pattern prediction model based on the pre-trained dataset to obtain the trained error pattern prediction model.

[0009] Furthermore, a pre-trained dataset is constructed, including: Generate a large number of numbers with length using a random number generator. The original binary data block; The target parameter is The BCH system code encoder encodes the original binary data block to obtain the corresponding codeword to be transmitted, where, The length of the BCH codeword. The length of the information bits. This indicates that a BCH codeword can correct a maximum of t random bit errors; The codeword to be transmitted is input into the channel model for modulation and noise superposition to simulate real communication channel conditions, obtain the corresponding BCH receive codeword, and generate BCH receive codeword sample data under different signal-to-noise ratio conditions to form a pre-training dataset.

[0010] Furthermore, the error pattern prediction model is pre-trained based on the pre-trained dataset, including: The BCH received codeword samples and corresponding real error patterns in the pre-trained dataset are used as training samples and input into the error pattern prediction model. Based on the difference between the bit-level error probability output by the error pattern prediction model and the corresponding real error pattern, a binary cross-entropy loss function is constructed and the model training loss is calculated. Based on the training loss, the Adam optimizer is used to update the network parameters of the error pattern prediction model, thereby optimizing the error pattern prediction capability for BCH received codewords. The prediction performance of the error pattern prediction model is evaluated during the pre-training process, and the model parameters are saved when the preset training termination condition is met, so as to obtain the trained error pattern prediction model.

[0011] Further, in step S2, the BCH received codeword to be decoded is input into the trained error pattern prediction model, and the predicted error pattern is output, including: The BCH received codeword to be decoded is input into the error pattern prediction model, and a forward inference operation is performed to obtain the bit-level error probability output corresponding to the length of the BCH received codeword. The bit-level error probability is compared with a preset decision threshold. When the error probability of the corresponding bit position is greater than or equal to the preset decision threshold, the bit position is determined to be erroneous, and the first decision value 1 is output. When the error probability of the corresponding bit position is less than the preset decision threshold, the bit position is determined not to be erroneous, and the second decision value 0 is output. Based on the decision results of each bit position, construct a length of... The binary error pattern, in which, The length of the BCH codeword is used to reflect the prediction error status of each bit position in the received BCH codeword.

[0012] Further, in step S2, the predicted error pattern is used to correct the BCH received codeword to be decoded, resulting in an estimated transmitted codeword, including: In the Galois domain GF(2), the BCH received codeword is XORed with the predicted error pattern to obtain the estimated transmitted codeword, and the relationship is expressed as: The estimated transmitted codeword is then stored in a buffer register for codeword validity determination. The BCH received codeword to be decoded. For the predicted error pattern, For the estimated transmitted codeword, is the bitwise XOR operator on the Galois field GF(2).

[0013] Further, step S3 includes: Perform syndrome calculation on the received BCH codeword to be decoded to obtain the syndrome. This is used to determine the validity of the BCH received codeword, where... The BCH received codeword to be decoded. For the parity check matrix corresponding to the BCH received codeword, the superscript... This is a transpose operation; When the syntactic expression of the BCH received codeword is 0, the BCH received codeword is a valid codeword, and the preceding part of the BCH received codeword is directly extracted. These bits serve as information bits in the BCH system code; When the syndrome of the BCH received codeword is not 0, the syndrome of the estimated transmitted codeword is calculated. When the syndrome of the transmitted codeword is 0, the estimated transmitted codeword is determined to be a valid codeword, and its first k bits are extracted as the information bits of the BCH system code. When the syndrome of the transmitted codeword is not 0, the error pattern prediction model is determined to have failed, and the traditional BCH decoding process is started to decode the BCH received codeword. At the same time, the error pattern prediction model is updated online based on the traditional decoding result. The traditional BCH decoding process is BM+Chien search.

[0014] Compared with the prior art, the present invention has at least one of the following technical advantages: This invention bypasses the bottleneck of serial iterative computation based on the Berlekamp-Massey algorithm and Chan's search algorithm in traditional BCH decoding by introducing an artificial intelligence model to predict error patterns of BCH received codewords in parallel. Error location determination can be completed in a single forward inference process, effectively reducing decoding latency and improving system throughput. At the same time, by constructing a dual-channel decoding control architecture based on synodal decision-making, the traditional BCH decoding path is introduced as a fallback verification method on the basis of the artificial intelligence prediction path. This improves the reliability of system error correction while ensuring decoding efficiency. When prediction fails, the traditional decoding results are used to perform online supervised updates to the model, enabling continuous evolution and better adapting to changes in the actual channel environment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below: Figure 1 This is a flowchart illustrating the steps of the parallel decoding method for BCH codes based on artificial intelligence according to the present invention. Figure 2 This is a flowchart of the BCH code decoding process in an embodiment of the present invention.

[0016] Figure 3 This is a diagram of the deep neural network architecture of the error pattern prediction model in an embodiment of the present invention.

[0017] Figure 4 This is a flowchart of the dual-channel decoding decision process in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0020] First Embodiment While existing BCH decoding technologies can meet basic communication requirements to a certain extent, they still suffer from performance bottlenecks when dealing with large-scale data transmission and complex channel conditions. The traditional Berlekamp-Massey (BM) algorithm and the Chien search algorithm are classic BCH decoding algorithms. They calculate error locations step-by-step through serial iteration, and this serial dependency limits the decoding speed. In high-throughput and low-latency communication scenarios, especially in applications with extremely high real-time requirements such as 5G and satellite communications, traditional serial decoding methods are insufficient. Although existing technologies attempt to improve decoding speed through hardware acceleration (such as parallelized Chien search) or algorithm optimization (such as the improved BM algorithm), these methods still cannot fundamentally overcome the bottleneck of serial computation, resulting in inefficiency when processing large-scale data. Therefore, existing technologies have significant limitations in improving decoding speed, reducing latency, and increasing throughput, necessitating a new decoding architecture to overcome this bottleneck.

[0021] To address the limitations of existing technologies, this application proposes an artificial intelligence-based parallel decoding method for BCH codes. This method transforms the traditional serial computation process into parallel computation by introducing a deep neural network model. Specifically, the AI ​​model learns the mapping relationship between received codewords and error patterns, enabling it to directly predict error patterns and bypassing the serial iterative steps in traditional decoding. This method significantly improves decoding speed, reduces decoding latency, and effectively increases throughput, making it particularly suitable for communication scenarios with extremely high latency and throughput requirements. Furthermore, this invention constructs a dual-channel decoding architecture, combining an AI path with a traditional verification path. The AI ​​decoding path accelerates the decoding process, while the traditional verification path serves as a guarantee for verification and error correction, ensuring the system's error correction reliability. During decoding, when the AI ​​model's prediction fails, the system can automatically switch to a traditional BCH decoder (such as the BM + Chien search algorithm) and perform online training and optimization of the AI ​​model based on the traditional decoding results. This dual-channel decoding control mechanism not only improves decoding efficiency but also enhances system reliability and the model's online evolution capability, enabling it to adapt to complex channel environments and communication conditions. Specific implementation details are as follows: like Figure 1 As shown, this invention provides a parallel decoding method for BCH codes based on artificial intelligence, including: Step S1: Obtain the BCH receive codeword to be decoded from the physical channel; Step S2: Input the BCH received codeword to be decoded into the trained error pattern prediction model, output the predicted error pattern, and use the predicted error pattern to correct the error of the BCH received codeword to be decoded to obtain the estimated transmitted codeword. Step S3: Perform synod calculations on the BCH received codeword to be decoded and the estimated transmitted codeword to determine the validity of the codeword; based on the results of the synod calculations, perform dual-channel decoding control in the order of first determining the BCH received codeword to be decoded and then determining the estimated transmitted codeword, and start the traditional BCH decoding process to supervise and update the model when the prediction fails, thereby improving the system reliability and model evolution capability.

[0022] Further, step S1 includes: The receiving antenna captures the modulated high-frequency analog signal from the physical channel and processes the high-frequency analog signal. The corresponding intermediate frequency analog signal is obtained by performing down-conversion and filtering. ; For intermediate frequency analog signals Perform analog-to-digital conversion to obtain a discrete-time series digital signal. and digital signals After carrier synchronization and symbol synchronization processing, the data is mapped onto the corresponding constellation diagram to obtain the corresponding binary code stream. Based on the frame synchronization information, a complete BCH codeword length data segment r is extracted from the binary bitstream as the BCH receive codeword to be decoded.

[0023] Furthermore, in step S2, the error pattern prediction model is a deep neural network architecture that has been pre-built and trained. The deep neural network architecture is based on the mapping relationship between the BCH received codewords and the corresponding error patterns learned in the pre-training stage.

[0024] Furthermore, through a pre-built and trained deep neural network architecture, the deep neural network architecture is based on the mapping relationship between the BCH received codewords and the corresponding error patterns learned in the pre-training stage, including: A deep neural network architecture for constructing an error pattern prediction model is proposed. The deep neural network structure consists of an input layer, at least three one-dimensional convolutional layers, at least two one-dimensional max-pooling layers, a flattening layer, a fully connected layer, and an output layer connected sequentially. The input layer receives data of dimension 1. A one-dimensional sequence, Let be the length of the BCH codeword, and 1 represent one input feature corresponding to each bit position. A one-dimensional sequence is input to a first-dimensional convolutional layer. After convolution processing, the output feature of the first-dimensional convolutional layer is passed to a first-dimensional max-pooling layer for downsampling. The output of the first-dimensional max-pooling layer serves as the input to a second-dimensional convolutional layer. The second-dimensional convolutional layer convolutions the pooled features and passes the output feature to the second-dimensional max-pooling layer for further downsampling. The output of the second-dimensional max-pooling layer is passed to a third-dimensional convolutional layer. After extracting high-level features, the convolution result is input to a flattening layer for dimensionality expansion. The flattened features are then input to a fully connected layer for feature fusion and comprehensive decision-making. The output of the fully connected layer is finally passed to the output layer, where an activation function generates a bit-level error probability corresponding to the BCH codeword length, indicating the probability of an error occurring at each bit position. Construct a pre-trained dataset and pre-train the error pattern prediction model based on the pre-trained dataset to obtain the trained error pattern prediction model.

[0025] Furthermore, a pre-trained dataset is constructed, including: Generate a large number of numbers with length using a random number generator. The original binary data block; The target parameter is The BCH system code encoder encodes the original binary data block to obtain the corresponding codeword to be transmitted, where, The length of the BCH codeword. The length of the information bits. This indicates that a BCH codeword can correct a maximum of t random bit errors; The codeword to be transmitted is input into the channel model for modulation and noise superposition to simulate real communication channel conditions, obtain the corresponding BCH receive codeword, and generate BCH receive codeword sample data under different signal-to-noise ratio conditions to form a pre-training dataset.

[0026] Furthermore, the error pattern prediction model is pre-trained based on the pre-trained dataset, including: The BCH received codeword samples and corresponding real error patterns in the pre-trained dataset are used as training samples and input into the error pattern prediction model. Based on the difference between the bit-level error probability output by the error pattern prediction model and the corresponding real error pattern, a binary cross-entropy loss function is constructed and the model training loss is calculated. Based on the training loss, the Adam optimizer is used to update the network parameters of the error pattern prediction model, thereby optimizing the error pattern prediction capability for BCH received codewords. The prediction performance of the error pattern prediction model is evaluated during the pre-training process, and the model parameters are saved when the preset training termination condition is met, so as to obtain the trained error pattern prediction model.

[0027] Specifically, such as Figure 3As shown, (1) a one-dimensional convolutional neural network is constructed, and its network architecture includes the following components: ① Input layer: receives a sequence of shape (n,1); ② Convolutional layer: uses multiple one-dimensional convolutional kernels to perform sliding convolution operations along the sequence direction. The initial layer uses a smaller kernel size to capture local features, and the deeper layers can use a larger kernel size to capture longer-range dependencies. The ReLU activation function is used after the convolutional layer to introduce nonlinearity; ③ Pooling layer: intermittently inserts one-dimensional max pooling layers to reduce the sequence length and the number of parameters; ④ Fully connected layer: flattens the feature map extracted by the convolutional layer and connects it to the fully connected layer for comprehensive judgment; ⑤ Output layer: a fully connected layer with n neurons and using the Sigmoid activation function. (2) Construction of the pre-training dataset, including the following sub-steps: ① Data generation: Use a random number generator to generate a large number of raw binary data blocks of length k; ② Encoding: Use the systematic code encoder of the target (n,k,t) BCH code to encode each data block to obtain the codeword to be transmitted; ③ Modulation and noise addition: Pass the codeword to be transmitted through an environment simulating a real channel to obtain the received codeword. Data is generated under different signal-to-noise ratio conditions to improve the generalization ability of the model. (3) Pre-training of the one-dimensional convolutional neural network model, including: ① Loss function: Use the binary cross-entropy loss function to measure the difference between the error probability distribution predicted by the model and the real label bit by bit; ② Optimizer: Use the Adam optimizer; ③ Training techniques: Use batch normalization, dropout method, and gradient pruning to prevent overfitting; ④ Evaluation metric: Monitor the bit error rate on an independent validation set to select the best model.

[0028] Further, in step S2, the BCH received codeword to be decoded is input into the trained error pattern prediction model, and the predicted error pattern is output, including: The BCH received codeword r to be decoded is input into the error pattern prediction model, and a forward inference operation is performed to obtain the bit-level error probability output corresponding to the length of the BCH received codeword. The bit-level error probability is compared with a preset decision threshold. When the error probability of the corresponding bit position is greater than or equal to the preset decision threshold, the bit position is determined to be erroneous, and the first decision value 1 is output. When the error probability of the corresponding bit position is less than the preset decision threshold, the bit position is determined not to be erroneous, and the second decision value 0 is output. Based on the decision results of each bit position, construct a length of... The binary error pattern, in which, The length of the BCH codeword is used to reflect the prediction error status of each bit position in the received BCH codeword.

[0029] Further, in step S2, the predicted error pattern is used to correct the BCH received codeword to be decoded, resulting in an estimated transmitted codeword, including: In the Galois domain GF(2), the BCH received codeword is XORed with the predicted error pattern to obtain the estimated transmitted codeword, and the relationship is expressed as: The estimated transmitted codeword is then stored in a buffer register for codeword validity determination. The BCH received codeword to be decoded. For the predicted error pattern, For the estimated transmitted codeword, is the bitwise XOR operator on the Galois field GF(2).

[0030] Further, step S3 includes: Perform syndrome calculation on the received BCH codeword to be decoded to obtain the syndrome. This is used to determine the validity of the BCH received codeword, where... The BCH received codeword to be decoded. For the parity check matrix corresponding to the BCH received codeword, the superscript... This is a transpose operation; When the syntactic expression of the BCH received codeword is 0, the BCH received codeword is a valid codeword, and the preceding part of the BCH received codeword is directly extracted. These bits serve as information bits in the BCH system code; When the syndrome of the BCH received codeword is not 0, the syndrome of the estimated transmitted codeword is calculated. When the syndrome of the transmitted codeword is 0, the estimated transmitted codeword is determined to be a valid codeword, and its first k bits are extracted as the information bits of the BCH system code. When the syndrome of the transmitted codeword is not 0, the error pattern prediction model is determined to have failed, and the traditional BCH decoding process is started to decode the BCH received codeword. At the same time, the error pattern prediction model is updated online based on the traditional decoding result. The traditional BCH decoding process is BM+Chien search.

[0031] Second Embodiment In this embodiment, in order to illustrate the technical effect of the artificial intelligence-based BCH code parallel decoding method, the following parameters were selected for experimental verification. However, it should be understood that the following parameters are only exemplary settings and do not constitute a limitation on the scope of protection of this invention.

[0032] Regarding the encoding parameters, a set of BCH codes is used, with a code length of n=1023, an information bit length of k=993, and an error correction capability of t=3, meaning that the BCH code can theoretically correct a maximum of 3 random bit errors. This parameter combination is one of the commonly used BCH code configurations in communication systems, facilitating comparative verification of the decoding performance of the method described in this invention.

[0033] In terms of the AI ​​model architecture, the error pattern prediction model adopts a one-dimensional convolutional neural network structure with an input dimension of (1, 1023, 1) to receive a BCH codeword sequence of length 1023. The network structure includes multiple levels of one-dimensional convolutional layers and pooling layers. The first convolutional layer has 32 kernels with a kernel size of 5, the second convolutional layer has 64 kernels with a kernel size of 3, and the third convolutional layer has 128 kernels with a kernel size of 3. Each convolutional layer uses the ReLU activation function. Max pooling layers with a pooling window size of 2 are set after some convolutional layers to reduce feature dimensionality and computational cost. The convolutional features are flattened and input to a fully connected layer with 256 neurons and the ReLU activation function. The output layer is a fully connected layer with 1023 neurons and the Sigmoid activation function, used to output the bit-level error probability corresponding to the BCH codeword length.

[0034] Regarding model training parameters, a dataset containing 1,000,000 samples was used in the pre-training phase, with 20% of the samples serving as the validation set. During training, binary cross-entropy was used as the loss function, and the Adam optimizer was used to update the model parameters. The learning rate was set to 0.0003, the number of training epochs was 50, and the batch size was 512. These training parameter settings enabled the model to fully learn the mapping relationship between BCH received codewords and corresponding error patterns.

[0035] In terms of channel simulation, an additive white Gaussian noise (AWGN) channel model was adopted, and the transmitted codeword was modulated using BPSK modulation. The signal-to-noise ratio was set to range from 2dB to 6dB to simulate the communication environment under different channel conditions, thereby verifying the decoding performance and generalization ability of the model under different noise levels.

[0036] During the online decoding stage, the decision threshold is set to 0.5 to perform binary decision on the bit-level error probability of the model output in order to generate the corresponding prediction error pattern.

[0037] The modular architecture diagram used in this embodiment is shown below. Figure 2 As shown, this embodiment specifically includes the following steps: S11, construct a one-dimensional convolutional neural network, such as Figure 3 As shown, its architecture design includes the following sub-steps: (1) Input layer: Receives a sequence of shape (1023,1); (2) Convolutional layers: Three one-dimensional convolutional kernels are used to perform sliding convolution operations along the sequence direction. The initial layer uses a smaller kernel size to capture local features, while deeper layers can use a larger kernel size to capture longer-range dependencies. A ReLU activation function is used after the convolutional layer to introduce non-linearity; (3) Pooling layer: Intermittently insert one-dimensional max pooling layers to reduce sequence length and the number of parameters; (4) Fully connected layer: The feature maps extracted by the convolutional layer are flattened and then connected to the fully connected layer for comprehensive judgment; (5) Output layer: A fully connected layer with 1023 neurons using the Sigmoid activation function.

[0038] S12, Pre-training dataset construction, includes the following sub-steps: (1) Data generation: A large number of raw binary data blocks of length 993 are generated using a random number generator; (2) Encoding: Each data block is encoded using a systematic code encoder for the target (1023, 993, 3) BCH code to obtain the codeword to be sent; (3) Modulation and noise addition: The codeword to be transmitted is passed through an AWGN channel to simulate the environment of a real channel and the received codeword is obtained. Data is generated under different signal-to-noise ratio conditions to improve the generalization ability of the model.

[0039] S13, Model Pre-training, includes the following elements: (1) Loss function: The binary cross-entropy loss function is used to measure the difference between the error probability distribution predicted by the model and the true label bit by bit; (2) Optimizer: Adam optimizer is used; (3) Training techniques: Use batch normalization, dropout method, and gradient clipping to prevent overfitting; (4) Evaluation metrics: Monitor bit error rate on independent validation sets to select the best model.

[0040] S14, Load the pre-trained model weight parameters; S21, the receiving antenna captures the modulated high-frequency analog signal from the physical channel. The signal is down-converted and filtered to obtain an intermediate frequency analog signal. ; S22, using an analog-to-digital converter to process analog signals Sampling is performed to obtain a discrete-time digital signal. ; S23, for digital signals Perform carrier synchronization and symbol synchronization, and map the signal onto the corresponding constellation diagram to obtain the binary code stream; S24. Based on the binary code stream and frame synchronization information, extract a complete BCH codeword length data segment r as the BCH receive codeword to be decoded. S31, input the received codeword r into the loaded neural network model to start a forward inference calculation; S32, the output probability value is compared with the decision threshold (set to 0.5). If the probability value is greater than or equal to the threshold, the bit position is determined to be faulty and 1 is output at the corresponding position; otherwise, it is determined to be fault-free and 0 is output at the corresponding position. S33, Format the binary sequence obtained after the above comparison into an error pattern of length n. , as the prediction error pattern output; S41, as Figure 4 As shown, in the Galois domain GF(2), the received codeword r is XORed with the predicted error pattern to obtain the estimated transmitted codeword, which is expressed as: S42, send the calculated estimate as a codeword Stored in the buffer register, awaiting further decision by the subsequent decision-maker; S51, the syndrome computation module performs parallel computation of the syndrome of codeword r. ; S52, if the syndrome calculation module outputs the syndrome... That is, r is a valid codeword, which does not need to be corrected. The information bits of the BCH systematic code can be recovered by directly extracting the first k bits of r. S53, if If the neural network model predicts the output of the estimated transmitted codeword It is a valid codeword (its syntactic form is...) ), then extract The first k bits are used as the information bits of the BCH system code; S54, if If the neural network model predicts the output of the estimated transmitted codeword It is not a valid codeword (its syntactic form is not 1). If r fails to predict, the neural network model will fail. In this case, a traditional decoder (BM+Chien search) is activated to decode r, improving system reliability and simultaneously performing online supervised training on the model to enhance its evolutionary capabilities.

[0041] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Many modifications and personalized choices of implementation methods can be made based on the content of this specification, including but not limited to different demodulation methods, different error correction and detection coding methods, and different segment arrangements. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A parallel decoding method for BCH codes based on artificial intelligence, characterized in that, include, Step S1: Obtain the BCH receive codeword to be decoded from the physical channel; Step S2: Input the BCH received codeword to be decoded into the trained error pattern prediction model, output the predicted error pattern, and use the predicted error pattern to correct the error of the BCH received codeword to be decoded to obtain the estimated transmitted codeword. Step S3: Perform synodal calculations on the BCH received codeword to be decoded and the estimated transmitted codeword to determine the validity of the codeword; based on the result of the synodal calculation, perform dual-channel decoding control in the order of first determining the BCH received codeword to be decoded and then determining the estimated transmitted codeword, and start the traditional BCH decoding process to supervise and update the model when the prediction fails.

2. The artificial intelligence-based parallel decoding method for BCH codes according to claim 1, characterized in that, Step S1 includes: The receiving antenna captures the modulated high-frequency analog signal from the physical channel, and performs down-conversion and filtering on the high-frequency analog signal to obtain the corresponding intermediate-frequency analog signal; The intermediate frequency analog signal is converted from analog to digital to obtain a discrete time series digital signal. After carrier synchronization and symbol synchronization processing, the digital signal is mapped onto the corresponding constellation diagram to obtain the corresponding binary code stream. Based on the frame synchronization information, a complete BCH codeword length data segment r is extracted from the binary code stream as the BCH received codeword to be decoded.

3. The artificial intelligence-based parallel decoding method for BCH codes according to claim 1, characterized in that, In step S2, the error pattern prediction model is a deep neural network architecture that is pre-built and trained. The deep neural network architecture is based on the mapping relationship between the BCH received codeword and the corresponding error pattern learned in the pre-training stage.

4. The artificial intelligence-based parallel decoding method for BCH codes according to claim 1, characterized in that, A pre-built and trained deep neural network architecture, based on the mapping relationship between BCH received codewords and corresponding error patterns learned during the pre-training phase, includes: The deep neural network architecture for constructing the error pattern prediction model comprises an input layer, at least three one-dimensional convolutional layers, at least two one-dimensional max-pooling layers, a flattening layer, a fully connected layer, and an output layer connected sequentially. The input layer receives data of dimension [missing information]. A one-dimensional sequence, The length of the BCH codeword is 1, and each bit position corresponds to one input feature. The one-dimensional sequence is input to a first-dimensional convolutional layer. After convolution processing, the first-dimensional convolutional layer passes the output feature to a first-dimensional max-pooling layer for downsampling. The output of the first-dimensional max-pooling layer is used as the input to a second-dimensional convolutional layer. The second-dimensional convolutional layer performs the convolution processing on the pooled features and passes the output feature to the second-dimensional max-pooling layer for further downsampling. The output of the second-dimensional max-pooling layer is passed to a third-dimensional convolutional layer. After extracting high-level features through the third-dimensional convolutional layer, the convolution result is input to the flattening layer for dimensionality expansion. The flattened features are input to the fully connected layer for feature fusion and comprehensive decision-making. The output of the fully connected layer is finally passed to the output layer, which generates a bit-level error probability corresponding to the length of the BCH codeword through an activation function, indicating the probability of an error occurring at each bit position. A pre-training dataset is constructed, and the error pattern prediction model is pre-trained based on the pre-training dataset to obtain the trained error pattern prediction model.

5. The artificial intelligence-based parallel decoding method for BCH codes according to claim 4, characterized in that, Constructing a pre-trained dataset includes: Generate a large number of numbers with length using a random number generator. The original binary data block; The target parameter is The BCH system code encoder encodes the original binary data block to obtain the corresponding codeword to be transmitted, wherein, The length of the BCH codeword. The length of the information bits. This indicates that the BCH codeword can correct a maximum of t random bit errors; The codeword to be transmitted is input into the channel model for modulation and noise superposition to simulate real communication channel conditions, thereby obtaining the corresponding BCH received codeword. Sample data of BCH received codewords are generated under different signal-to-noise ratio conditions to form the pre-training dataset.

6. The artificial intelligence-based parallel decoding method for BCH codes according to claim 4, characterized in that, Pre-training the error pattern prediction model based on the pre-trained dataset includes: The BCH received codeword samples and corresponding real error patterns in the pre-trained dataset are used as training samples and input into the error pattern prediction model. Based on the difference between the bit-level error probability output by the error pattern prediction model and the corresponding real error pattern, a binary cross-entropy loss function is constructed and the model training loss is calculated. Based on the training loss, the Adam optimizer is used to update the network parameters of the error pattern prediction model, thereby optimizing the error pattern prediction capability for the BCH received codeword. The prediction performance of the error pattern prediction model is evaluated during the pre-training process, and the model parameters after training are saved when the preset training termination condition is met, so as to obtain the trained error pattern prediction model.

7. The artificial intelligence-based parallel decoding method for BCH codes according to claim 6, characterized in that, In step S2, the BCH received codeword to be decoded is input into the trained error pattern prediction model, and the predicted error pattern is output, including: The BCH received codeword to be decoded is input into the error pattern prediction model, and a forward inference operation is performed to obtain the bit-level error probability output corresponding to the length of the BCH received codeword; The bit-level error probability is compared with a preset decision threshold. When the error probability corresponding to the bit position is greater than or equal to the preset decision threshold, it is determined that an error has occurred at the bit position, and a first decision value of 1 is output. When the error probability corresponding to the bit position is less than the preset decision threshold, it is determined that no error has occurred at the bit position, and a second decision value of 0 is output. Based on the decision results of each bit position, construct a length of... The binary error pattern, wherein, The length of the BCH codeword is given, and the error pattern is used to reflect the prediction error state of each bit position in the received BCH codeword.

8. The artificial intelligence-based parallel decoding method for BCH codes according to claim 7, characterized in that, In step S2, the predicted error pattern is used to correct errors in the BCH received codeword to be decoded, resulting in an estimated transmitted codeword, including: On the Galois domain GF(2), the BCH received codeword is XORed with the predicted error pattern to obtain the estimated transmitted codeword, and the relationship is expressed as: The estimated transmitted codeword is then stored in a buffer register for validity determination. The BCH received codeword to be decoded. For the predicted error pattern, For the estimated transmitted codeword, Let XOR be the bitwise XOR operator on the Galois field GF(2).

9. The artificial intelligence-based parallel decoding method for BCH codes according to claim 8, characterized in that, Step S3 includes: Perform the syndrome calculation on the BCH received codeword to be decoded to obtain the syndrome. This is used to determine the validity of the BCH received codeword, wherein, The BCH received codeword to be decoded. For the parity check matrix corresponding to the BCH received codeword, the superscript... This is a transpose operation; When the syndrome of the BCH received codeword is 0, the BCH received codeword is a valid codeword, and the preceding part of the BCH received codeword is directly extracted. These bits serve as information bits in the BCH system code; When the syndrome of the BCH received codeword is not 0, the syndrome of the estimated transmitted codeword is calculated. When the syndrome of the transmitted codeword is 0, the estimated transmitted codeword is determined to be a valid codeword, and its first k bits are extracted as information bits of the BCH system code. When the syndrome of the transmitted codeword is not 0, the error pattern prediction model is determined to have failed, and the traditional BCH decoding process is started to decode the BCH received codeword. At the same time, the error pattern prediction model is updated online under supervision based on the traditional decoding result. The traditional BCH decoding process is BM+Chien search.

Images