Intelligent Turbo channel decoding method based on model and data dual drive

Abstract

The invention discloses an intelligent Turbo channel decoding method based on model and data dual drive, and belongs to the field of channel decoding. The method comprises the steps of generating a training data set; an APETurbo decoding model based on dual driving of the model and the data is constructed; training the APETurbo decoding model by using the training data set, and storing model weight parameters; and S3, loading the model weight parameters stored in the step S3, and deploying the APETurbo decoding model at a receiving end of a communication system for channel decoding. Compared with a traditional Turbo decoding algorithm, the method has the advantage that the bit error rate is remarkably reduced. Compared with an existing Turbo decoding algorithm based on model driving, traditional linear calculation is reserved, nonlinear feature extraction is added, the algorithm interpretability is kept while the nonlinear feature modeling capacity is improved, and the decoding accuracy of the model is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of channel decoding, and particularly relates to an intelligent Turbo channel decoding method based on dual driving of model and data. Background Art

[0002] Under the background of the large-scale commercialization of the fifth-generation mobile communication (5G) technology and the pre-research of the sixth-generation mobile communication (6G) standard, the communication system is facing multiple technical challenges in terms of transmission rate, reliability, and complex scenario adaptability. The demand for high-efficient and reliable information transmission in high-end fields such as deep space communication, satellite communication, military communication, and UAV communication is increasing day by day. The traditional Turbo code has excellent error correction ability and can approach the Shannon limit performance. As a classic channel coding scheme, it shows excellent error correction performance in the above-mentioned various communication scenarios.

[0003] Channel decoding refers to the process of making decisions on the received symbol messages. The main research objective of channel decoding is to improve the reliability of information transmission and ensure that the receiving end can accurately restore the information sent by the sending end. Through channel decoding, the receiving end can use the redundant information added during encoding for error detection and correction, thereby reducing the bit error rate and improving the performance of the communication system. Channel decoding is a key link in the communication system to ensure the reliability and efficiency of information transmission, and the optimization of its performance is crucial for improving the performance of the entire communication system.

[0004] The core architecture of the Turbo code decoding technology is designed based on the Parallel Concatenated Convolutional Code (PCCC). Its core feature lies in realizing the iterative soft information interaction mechanism of the Recursive Systematic Convolutional Code (RSC) component decoders through an interleaver. During the decoding process, the receiving end parallel inputs the soft information sequence output by the channel into two Soft-Input Soft-Output (SISO) component decoders. Each component decoder calculates the log-likelihood ratio of the posterior probability based on the received sequence and the prior information from the other decoder, and finally generates the Extrinsic Information. After being interleaved or de-interleaved, this extrinsic information is input into the other component decoder as the updated prior information to form a closed-loop iterative feedback. Through multiple iterations, the statistical characteristics of the extrinsic information tend to converge. At this time, a hard decision is made on the posterior log-likelihood ratio (Posteriori LLR), and the optimal estimated sequence of the information bits can be obtained.

[0005] Turbo codes rely on precise mathematical models and a large number of iterative calculations to achieve efficient error correction performance and have been widely used in fields such as wireless communication. Currently, the mainstream Turbo decoding implementation schemes mainly fall into two categories: one is based on the Maximum A Posteriori (MAP) algorithm and its improved algorithms (such as the Max-Log-MAP algorithm). The other is the Soft-Output Viterbi Algorithm (SOVA) and its improved algorithms. The MAP algorithm calculates the trellis path metric through forward-backward recursion to accurately solve the bit-level posterior probability. However, its computational complexity grows exponentially with the constraint length and code length, making it difficult to meet the requirements of high-real-time scenarios. To reduce the complexity, the Max-Log-MAP algorithm uses logarithmic domain approximation operations to compress the complexity to a linear order, but introduces a certain performance loss. The SOVA algorithm achieves fast decoding through path metric pruning, but its soft output accuracy is significantly lower than that of the MAP class algorithms.

[0006] Traditional Turbo code decoding algorithms (such as Max-Log-MAP) rely on the accurate transmission of prior information to complete decoding during the iterative process. However, in actual scenarios, there are various noise interferences and nonlinear effects, resulting in poor quality of the initial channel soft information input to the decoder. When calculating the posterior probability log-likelihood ratio of the decoder output, it is also affected by noise, leading to inaccurate posterior probability estimation. In the subsequent iterative decoding process, the errors of the prior information and the channel estimation deviation will accumulate with the number of iterations, resulting in performance degradation, making it difficult to face complex and changing communication environments and channel conditions, and affecting its application in actual scenarios. To compensate for the performance loss, the number of iterations needs to be increased (such as >10 times), resulting in a linear increase in the decoding delay.

[0007] Deep learning is an important research direction in the field of Artificial Intelligence (AI) and has demonstrated excellent performance in various application scenarios, including image recognition, speech recognition, natural language processing, etc. Deep learning performs data representation learning through multi-layer neural networks, can automatically extract features from a large amount of data, and achieve the recognition and modeling of complex patterns, and has demonstrated its powerful capabilities and potential in multiple application scenarios. In the future, with the continuous in-depth research and the continuous development of technology, deep learning is expected to make breakthroughs in more fields.

[0008] Deep learning (DL) can perform data representation learning through a multi-layer neural network, automatically extract features from a large amount of data, and achieve the recognition and modeling of complex patterns. It has been widely applied in computer vision, natural language processing, etc., and has gradually penetrated into various research fields. In recent years, the combination of deep learning and communication technology has also attracted extensive attention. Deep learning technology centered on the deep neural network (DNN), with its powerful non-linear modeling and feature self-learning capabilities, shows great potential in the field of channel decoding, providing new ideas and technical paths for breaking through the performance bottleneck of traditional algorithms. However, existing decoding schemes based on deep learning still face some technical challenges. Building an intelligent decoding architecture with higher decoding accuracy, interpretability, and compatibility with existing communication standards has become one of the technical challenges for promoting the evolution of the communication physical layer to 6G. Summary of the Invention

[0009] In order to overcome the problem of limited bit error rate performance of existing Turbo decoding methods, the present invention proposes an intelligent Turbo channel decoding method based on dual drive of model and data. This method significantly improves the bit error rate performance of decoding by designing learnable neural network components, residual connection structures, and hybrid loss functions.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] An intelligent Turbo channel decoding method based on dual drive of model and data, comprising the following steps:

[0012] S1: Generate a training data set;

[0013] S2: Construct an APETurbo decoding model based on dual drive of model and data;

[0014] S3: Use the training data set to train the APETurbo decoding model and save the model weight parameters;

[0015] S4: Load the model weight parameters saved in step S3 and deploy the APETurbo decoding model at the receiving end of the communication system for channel decoding.

[0016] Further, the specific manner of step S1 is:

[0017] S11: Build a Turbo channel encoding and decoding transmission system in the Pycharm simulation platform. The Turbo channel encoding and decoding transmission system includes a sending end, an AWGN channel, and a receiving end;

[0018] S12: Generate an information bit sequence at the transmitter of the Turbo channel coding and transmission system; then, perform Turbo coding on the information sequence using a parallel concatenated convolutional code structure to generate a coded sequence;

[0019] S13: Modulate the coded sequence, transmit the information bits through an AWGN channel, add Gaussian white noise in the channel, and send it to the receiver;

[0020] S14: The receiver receives the information sequence with added Gaussian white noise, demodulates the information sequence, calculates the channel soft output, decodes it using the Log-MAP algorithm, and obtains the log-likelihood ratio of the posterior probability of the signal output in the last iteration;

[0021] S15: Combine the channel soft output obtained at the receiver, the information bits sent by the transmitter, and the log-likelihood ratio of the posterior probability obtained in step S14 into a set of data as a training data;

[0022] S16: Collect multiple training data to obtain a training data set.

[0023] Furthermore, the APE Turbo decoding model includes one or more intelligent decoding units with the same structure and an auxiliary hard decision maker. Each intelligent decoding unit includes two identical interleavers, two identical SISO decoders, a deinterleaver corresponding to the interleaver, and two APENet networks with the same structure but different weight parameters;

[0024] The APENet network has 6 layers: the first layer is the input layer, the middle 4 layers are hidden layers, and the last layer is the output layer; where:

[0025] The input layer consists of 3K neurons, representing the log-likelihood ratio information L(u k |y) of the posterior probability of K bits, system information and prior information L(u k );

[0026] The first hidden layer is a linear part calculation layer, and the expression is as follows:

[0027]

[0028] Among them, L e (u k ) is the extrinsic information obtained by linearly calculating through learnable parameters, L(u k |y) is the log-likelihood ratio information of the posterior probability output by the SISO decoder, is the system information, L(u k) is the prior information. Define w1, w2, and w3 as the learnable weights for the three inputs acting on the first hidden layer; the output data shape of the first hidden layer is 1×K;

[0029] The second hidden layer is a data flattening layer, and the output data shape of the second hidden layer is K×1;

[0030] The third and fourth hidden layers are non-linear feature extraction layers: among them, the third hidden layer uses a fully connected layer with an input of 1 dimension and an output of 64-dimensional features, the weight matrix shape is 64×1, the output shape is K×64, and the PReLU non-linear activation function is used after the fully connected operation; the fourth hidden layer uses a fully connected layer with an input dimension of 64 and an output of 1 dimension, followed by the Tanh non-linear activation function, and the output is the non-linear feature extraction term of the external information obtained by neural network processing;

[0031] The output layer restores the original data dimension of 1×K through data reshaping, constructs a residual connection structure, and mixes the calculation result of the linear combination with the learnable parameter and the non-linear feature extraction term of the external information obtained by neural network processing. The expression is as follows:

[0032] L e_NN (u k ) = L e (u k ) + α·L e_NNout (u k )

[0033] Among them, α is the learnable mixing coefficient, α∈[0,1], L e (u k ) represents the external information obtained through linear calculation, L e_NNout (u k ) represents the non-linear feature extraction term obtained by enhancing the features of L e (u k ) with a neural network, and L e_NN (u k ) represents the corrected external information obtained by combining the linear calculation part and the non-linear feature extraction part in a certain proportion.

[0034] Furthermore, the input of the APETurbo decoding model is system information the first parity check information the second parity check information and the prior information Among them, the system information, the first parity check information, and the second parity check information are respectively input into each intelligent decoding unit. The input to the first intelligent decoding unit is initialized to 0;

[0035] Suppose there are a total of M intelligent decoding units, and each intelligent decoding unit includes a first SISO decoder, a second SISO decoder, a first APENet network, and a second APENet network;

[0036] In the first intelligent decoding unit, the received and are sent to the first SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and are sent to the first APENet network together; the result output by the first APENet network is the more accurately estimated extrinsic information After interleaving, the prior information input to the second SISO decoder is obtained The and the interleaved systematic information are sent to the second SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved systematic information and are sent to the second APENet network together; the result output by the second APENet network is the more accurately estimated extrinsic information After deinterleaving, the prior information required for input to the first SISO decoder of the next intelligent decoding unit is obtained

[0037] In the subsequent m-th intelligent decoding unit, m = 2, 3,..., M - 1, the received and the prior information output by the previous intelligent decoding unit are sent to the first SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and are sent to the first APENet network together; the result output by the first APENet network is the more accurately estimated extrinsic information After interleaving, the prior information input to the second SISO decoder is obtained The and the interleaved systematic information are sent to the second SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved systematic information and are sent to the second APENet network together; the result output by the second APENet network is the more accurately estimated extrinsic information After deinterleaving, the prior information required for input to the first SISO decoder of the next intelligent decoding unit is obtained

[0038] In the Mth intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent to the first SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and which are sent to the first APENet network together; the result output by the first APENet network is the more accurately estimated extrinsic information which, after interleaving, becomes the prior information input to the second SISO decoder The and the interleaved systematic information are sent to the second SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved systematic information and are sent to the second APENet network together; the result output by the second APENet network is the more accurately estimated extrinsic information which, after deinterleaving, is added to the systematic information to obtain the log-likelihood ratio L M (u k │y) of the posterior probability of each information bit in the original information sequence. The log-likelihood ratio of the posterior probability of each information bit is normalized using the Sigmoid function, and the output is o k For o k a hard decision is made to obtain the final decoded sequence

[0039] Furthermore, the weight parameters of the APENet network include: three weights w1, w2, w3 acting on the log-likelihood ratio information L(u k |y) of the posterior probability output by the SISO decoder, the systematic information the prior information L(u k ); the weights and biases of the fully connected layer; the learnable parameter a in the PReLU non-linear activation function; the learnable mixing coefficient α of the output layer, where α ∈ [0, 1].

[0040] Furthermore, the training process of the APETurbo decoding model is as follows:

[0041] (1) Demultiplex the channel soft output in the training dataset to obtain the systematic information the parity-check information and the parity-check information The and the prior information initialized to 0 are input into the APETurbo decoding model for decoding;

[0042] (2) In each training cycle, the APETurbo decoding model processes the input data and outputs the log-likelihood ratio of the posterior probability of each bit through the last intelligent decoding unit; the log-likelihood ratio of the posterior probability of a batch of data is used as the model prediction value, and the log-likelihood ratio of the posterior probability of each bit obtained by decoding the batch of data in the training dataset using the Log-MAP algorithm is used as the model true value;

[0043] (3) Normalize the prediction value and the true value using the Sigmoid function, calculate the mean squared error loss of all data in a batch, sum and take the average; perform regularization constraint on the mixing coefficient α to obtain the regularization constraint term; define a mixed loss function to train and update the parameters of the APETurbo network model; the expression of the mixed loss function is as follows:

[0044]

[0045] where, is the average mean squared error loss function of a batch of data, σ(·) represents the Sigmoid function, which maps the log-likelihood ratio to the probability space; assume the training batch size is B, that is, there are B message sequences u of length K (1) , u (2) , …, u (B) , u (i) ∈ {0, 1} K , i = 1, 2, ..., B, then y (1) , y (2) , …, y (B) is the log-likelihood ratio information of the received information bits; is the log-likelihood ratio prediction value of the posterior probability of each bit in each training data; is the true value; is the regularization constraint on the mixing coefficient α, λ is the regularization strength hyperparameter, which is used to constrain the numerical stability of the learnable mixing coefficient α, θ j represents the j-th learnable mixing coefficient α in the entire network;

[0046] (4) Observe the loss function. When the model converges or reaches the maximum number of training times, the training is completed and the weight parameters of the model are saved.

[0047] Further, the specific manner of step S4 is as follows:

[0048] S41: Load the saved model parameters and deploy the APETurbo decoding model at the receiving end of the communication system;

[0049] S42: The receiving end receives the noisy information sequence transmitted by the sending end, demodulates the noisy information sequence, and calculates the channel soft output;

[0050] S43: Demultiplex the soft output of the channel to obtain system information, the first parity information, and the second parity information, and input this information together with the a priori information initialized to 0 into the APETurbo decoding model for channel decoding to obtain the decoded information sequence.

[0051] The beneficial effects of the present invention are as follows:

[0052] (1) Compared with the traditional Turbo channel decoding algorithm, the present invention reconstructs the iterative decoding process into a multi-level cascaded neural network architecture through a deep unfolding mechanism, and introduces trainable parameters and lightweight neural network modules for optimization. Under the AWGN channel and the same number of iterations, the bit error rate performance is improved compared with the traditional Turbo channel decoding algorithm.

[0053] (2) Different from the existing model-driven Turbo decoders, the present invention constructs a lightweight feature enhancement module including a fully connected layer and a non-linear activation function (PReLU / Tanh) in the APENet neural sub-network to capture complex channel characteristics, improving the model's ability to extract channel non-linear features and the model's ability to model complex channel responses. Compared with the existing model-driven intelligent Turbo decoders, the bit error rate performance is improved under the AWGN channel and the same number of iterations.

[0054] (3) Compared with the existing intelligent Turbo decoders, the present invention adopts a residual connection structure, and realizes the dual-path fusion by adopting a learnable mixing coefficient α (α ∈ [0, 1]) that balances the original linear output and the non-linear extraction term enhanced by the neural network. While retaining the interpretability of traditional linear calculations, the model's ability to extract channel non-linear features is improved through gated non-linear correction, and the decoding accuracy of the model is improved.

[0055] (4) The present invention constructs a hybrid loss function including a mean square error term in the probability space and a regularization constraint, and jointly optimizes the decoding accuracy and parameter stability. By minimizing the mean square error between the predicted LLR and the true LLR in the probability space and suppressing the overfitting tendency of the parameter α, overfitting caused by the excessive proportion of the non-linear enhancement path is prevented, and the generalization ability and decoding accuracy of the model are improved. Description of the Drawings

[0056] Figure 1 is the flowchart of the embodiment of the present invention.

[0057] Figure 2 is the schematic diagram of the principle of the APETurbo decoder.

[0058] Figure 3 is the schematic diagram of the structure of the intelligent decoding unit.

[0059] Figure 4 It is a schematic diagram of the principle of the APENet neural sub-network model.

[0060] Figure 5 It is the decoding flowchart of the communication system deployed with the APETurbo decoding model in the embodiment of the present invention.

[0061] Figure 6 It is the performance simulation result diagram of the method in the embodiment of the present invention. Detailed implementation manners

[0062] The present invention will be described in detail below with reference to the accompanying drawings.

[0063] An intelligent Turbo channel decoding method based on model and data double drive, as Figure 1 shown, includes the following steps:

[0064] S1: Generate a training data set;

[0065] S2: Construct an APETurbo decoding model based on model and data double drive;

[0066] S3: Use the training data set to train the APETurbo decoding model; observe the loss function to ensure good convergence of the model, and save the model weight parameters;

[0067] S4: Load the model weight parameters saved in step S3, and deploy the APETurbo decoding model at the receiving end of the communication system for channel decoding.

[0068] Further, the specific steps of step S1 include:

[0069] S11: Build a Turbo channel encoding and decoding transmission system in the Pycharm simulation platform, and the Turbo channel encoding and decoding transmission system includes a sending end, an AWGN channel, and a receiving end;

[0070] S12: Generate an information bit sequence through the sending end of the Turbo channel encoding and decoding transmission system; then, perform Turbo encoding on the information sequence using a parallel concatenated convolutional code structure to generate an encoded sequence;

[0071] S13: Modulate the encoded sequence and transmit it through the AWGN channel, add Gaussian white noise with a specified signal-to-noise ratio, and transmit it to the receiving end;

[0072] S14: The receiving end receives the noisy information sequence, demodulates it, calculates the channel soft output, decodes it using the Log-MAP algorithm, and obtains the logarithmic likelihood ratio of the posterior probability of the signal output in the last iteration;

[0073] S15: Combine the channel soft output obtained at the receiving end, the information bits transmitted by the transmitting end, and the log-likelihood ratio of the posterior probability of the information bits obtained by the Log-MAP decoding algorithm into a set of data as a training data; among them, demultiplexing the channel soft output will obtain the system information input to the APETurbo decoder Parity information And parity information The information bits transmitted by the transmitting end are used to calculate the bit error rate of the training set at the end of each training step; the log-likelihood ratio of the posterior probability obtained by the Log-MAP algorithm decoding is used as the training objective for optimizing the APETurbo decoding model;

[0074] S16: Collect multiple training data to obtain a training data set.

[0075] As Figure 2 shown, the APETurbo decoding model reconstructs the M - iteration process of traditional Turbo channel decoding into an M - layer cascaded neural network architecture through deep unfolding. Generally, M is taken as 3, that is, 3 iterations. Each iteration level is represented by an intelligent decoding unit. The main structure includes: one or more intelligent decoding units with the same structure and an auxiliary hard decision maker. Each intelligent decoding unit includes two identical interleavers, two identical SISO decoders, a de - interleaver corresponding to the interleaver, and two APENet networks with the same structure but different weight parameters. In each intelligent decoding unit, the modules of the two SISO decoders and their corresponding APENet neural sub - networks are respectively used to calculate the extrinsic information of the normal layer and the interleaved layer.

[0076] Furthermore, the APENet neural sub - network model is composed of a linear part calculation module and a prior information non - linear feature extraction module, with a total of 6 layers: the first layer is the input layer, the middle 4 layers are hidden layers, and the last layer is the output layer. The specific structure of each layer of the APENet neural sub - network is as follows:

[0077] K represents the length of the information sequence. The input layer consists of 3K neurons, and the log - likelihood ratio of the posterior probability information L(u k |y), system information and prior information L(u k ).

[0078] The first hidden layer is a linear part calculation layer: Define a triple learnable weight matrix, which acts on the three inputs of the APENet neural sub - network respectively. Then, through bit - wise weight sharing, the triple learnable weight matrix obtains three learnable weights w1, w2, w3 that need to be trained in the first layer of the neural sub - network model. The output of hidden layer 1 is the extrinsic information obtained by linearly combining and calculating after adjusting the learnable parameters. The specific expression is as follows:

[0079]

[0080] Among them, L e (u k ) is the external information obtained by linearly calculating through learnable parameter adjustment. L(u k |y) is the log-likelihood ratio information of the posterior probability output by the SISO decoder. is the system information. L(u k ) is the prior information. Define w1, w2, and w3 as the learnable weights acting on the three inputs of the first hidden layer; the output data shape of the first hidden layer is 1×K;

[0081] The second hidden layer is a data flattening layer, and the output data shape of the second hidden layer is K×1;

[0082] The third and fourth hidden layers are non-linear feature extraction layers: among them, the third hidden layer uses a fully connected layer with an input of 1 dimension and an output of 64-dimensional features. The weight matrix shape is 64×1, and the output shape is K×64. After the fully connected operation, the PReLU non-linear activation function is used; the fourth hidden layer uses a fully connected layer with an input dimension of 64 and an output of 1 dimension, followed by the Tanh non-linear activation function, and the output is the non-linear feature extraction term of the external information obtained through neural network processing;

[0083] The output layer restores the original data dimension of 1×K through data reshaping, constructs a residual connection structure, mixes the calculation result of the linear combination with the learnable parameter and the non-linear feature extraction term of the external information obtained through neural network processing, uses the output of the linear part calculation layer as the reference path, and uses the non-linear correction term output by the prior information non-linear feature enhancement module as the residual path to achieve gradient stability and feature enhancement through skip connection. The expression is as follows:

[0084] L e_NN (u k ) = L e (u k ) + α·L e_NNout (u k )

[0085] Among them, α is the learnable mixing coefficient, α ∈ [0, 1], initialized to 0.5. L e (u k ) represents the external information obtained through linear calculation. L e_NNout (u k ) represents the non-linear feature extraction term obtained by enhancing the features of L e (u k ) using a neural network. L e_NN (u k)It represents the corrected external information obtained by combining the linear calculation part and the non - linear feature extraction part in a certain proportion.

[0086] The learnable mixing coefficient α is used to balance the original linear calculation output value and the non - linear feature extraction value enhanced by the neural network, dynamically adjust the fusion ratio of the two, and achieve the adaptive mixing of the two paths.

[0087] Refer to Figure 3 , in each iteration of the APETurbo decoding model, first, the data passes through the SISO decoder to call the Max - Log - MAP algorithm to calculate the external information. Subsequently, the external information is processed by the APENet neural sub - network to obtain a more accurately estimated external information, and the interleaving / de - interleaving operation is performed to obtain the prior information for the interaction between decoders, realizing the information interaction between the two decoders. Suppose there are a total of M intelligent decoding units, and the specific decoding process includes:

[0088] In the first intelligent decoding unit, the received and are sent to the SISO decoder D1 for decoding to obtain the posterior probability log - likelihood ratio information of each information bit and are sent to the APENet neural sub - network 1 together; the result output by sub - network 1 is a more accurately estimated external information After interleaving, it becomes the prior information input to the SISO decoder D2 The and the interleaved systematic information are sent to the SISO decoder 2 for decoding to obtain the posterior probability log - likelihood ratio information of each information bit in the interleaved sequence The interleaved systematic information and are sent to the APENet neural sub - network 2 together; the result output by sub - network 2 is a more accurately estimated external information After de - interleaving, it becomes the prior information required for the input SISO decoder D1 of the next intelligent decoding unit

[0089] In the subsequent m - th intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent to the SISO decoder D1 for decoding to obtain the posterior probability log - likelihood ratio information of each information bit and are sent to the APENet neural sub - network m together; the result output by sub - network m is a more accurately estimated external information After interleaving, it becomes the prior information input to the SISO decoder D2 The Together with the interleaved system information, it is sent to the SISO decoder D2 for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved system information and are sent to the APENet neural sub-network 2m together; the result output by the sub-network 2m is the more accurately estimated extrinsic information After deinterleaving, the prior information required for input to the SISO decoder D1 of the next intelligent decoding unit is obtained

[0090] In the Mth intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent to the SISO decoder D1 for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and are sent to the APENet neural sub-network M together; the result output by the sub-network M is the more accurately estimated extrinsic information After interleaving, the prior information for input to the SISO decoder D2 is obtained The and the interleaved system information are sent to the SISO decoder D2 for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence After deinterleaving, the log-likelihood ratio L M (u k │y) of the posterior probability of each information bit is obtained. The log-likelihood ratio of the posterior probability of each information bit is normalized using the Sigmoid function, and o k is output. Hard decision is performed on o k to obtain the final decoded sequence Furthermore, each iteration level of the constructed complete APETurbo decoder contains two APENet neural sub-network models, which are used to extract the non-linear features of the extrinsic information, and the extrinsic information calculated by the traditional linear combination is combined with the non-linear correction term of the extrinsic information enhanced by the neural network through residual connection. The output of the model is the corrected extrinsic information

[0091] The weight parameters of the APENet network include: the three weights w1, w2, w3 acting on the log-likelihood ratio information L(u k |y) of the posterior probability output by the SISO decoder, the system information the prior information L(u k ); the weights and biases of the fully connected layer; the learnable parameter a in the PReLU non-linear activation function; the learnable mixing coefficient α of the output layer, where α ∈ [0, 1]

[0092] The training process of the APE Turbo decoding model is as follows:

[0093] (1) Input the in the training dataset and the prior information initialized to 0 into the APE Turbo decoding model for decoding;

[0094] (2) In each training cycle, the APE Turbo decoding model processes the input data to obtain the log-likelihood ratio of the posterior probability of each bit output by the last intelligent decoding unit of each data. The log-likelihood ratio of the posterior probability of a batch of data is used as the model prediction value; the log-likelihood ratio of the posterior probability of each bit obtained by decoding the batch of data using the Log-MAP algorithm is used as the true value;

[0095] (3) Normalize the prediction value and the true value using the Sigmoid function, calculate the mean square error loss of all data in a batch, sum and take the average; perform regularization constraints on the mixing coefficient α to obtain the regularization constraint term; define a mixed loss function to train and update the parameters of the APE Turbo network model. The loss function is a mixed loss function, which optimizes the decoding accuracy and parameter stability simultaneously through a dual supervision mechanism. Its mathematical structure includes a core loss term and a regularization constraint, and its expression is as follows:

[0096]

[0097] Among them, the first part uses the Mean Square Error (MSE) loss function, σ(·) represents the Sigmoid function, which maps the log-likelihood ratio to the probability space; assume that the training batch size is B, that is, there are B message sequences u (1) , u (2) , …, u (B) , u (i) ∈{0, 1} K , i = {1, 2, …, B}, then y (1) , y (2) , …, y (B) is the log-likelihood ratio information of the received information bits, is the predicted value of the log-likelihood ratio of the posterior probability of each bit in each training data; is the true value. The second part is the regularization constraint on the mixing coefficient α. λ is the regularization strength hyperparameter, which is used to constrain the numerical stability of the learnable mixing coefficient α. θ j represents the j-th learnable mixing coefficient α in the entire network. This loss function realizes joint optimization by minimizing the mean square error between the predicted LLR and the true LLR in the probability space and suppressing the overfitting tendency of the parameter α.

[0098] (4) Observe the loss function to ensure that the model converges well and save the model weight parameters. When the number of training epochs reaches the set number, the training ends; otherwise, return to step 1 to continue training.

[0099] As Figure 5 shown, the specific method of step S4 is as follows:

[0100] S41: Load the saved model parameters and deploy the APETurbo decoding model at the receiving end of the communication system;

[0101] S42: The receiving end receives the noisy information sequence transmitted from the sending end, demodulates the noisy information sequence, and calculates the channel soft output;

[0102] S43: Demultiplex the channel soft output to obtain the system information, the first parity-check information, and the second parity-check information. Input these information together with the prior information initialized to 0 into the APETurbo decoding model for channel decoding to obtain the decoded information sequence.

[0103] In a simulation environment, the performance of the APETurbo decoding model can be tested. The specific process is as follows:

[0104] Build a Turbo channel coding and decoding transmission system in the Pycharm simulation platform. The Turbo channel coding and decoding transmission system includes a sending end, an AWGN channel, and a receiving end;

[0105] Generate an information bit sequence through the sending end of the Turbo channel coding and decoding transmission system; then, perform Turbo coding on the information sequence using a parallel concatenated convolutional code structure to generate a coded sequence;

[0106] Modulate the coded sequence and transmit it through the AWGN channel, adding Gaussian white noise with a specified signal-to-noise ratio, and send it to the receiving end;

[0107] The receiving end receives the noisy information sequence, demodulates it, and calculates the channel soft output;

[0108] Demultiplex the channel soft output to obtain the system information input to the APETurbo decoder parity-check information and parity-check information together with the information bits sent by the sending end to form a set of data, which is used as test data in a specific signal-to-noise ratio environment;

[0109] Collect multiple training data to obtain a training data set in a specific signal-to-noise ratio environment; collect training data sets in multiple signal-to-noise ratio environments;

[0110] The system information in the training data set collected in a specific signal-to-noise ratio environment parity-check information Parity check information Input the APETurbo decoding model, and initialize the prior information input to the model to 0; obtain the decoded sequence through the processing of the APETurbo decoding model; calculate the bit error rate of the test data set of the information bits sent by the transmitter under each signal-to-noise ratio environment according to the information bits sent by the transmitter, and obtain the bit error rate performance curve.

[0111] The following is a more specific example:

[0112] An intelligent Turbo channel decoding method based on dual driving of model and data, whose decoding architecture is a cascaded neural network with deep unfolding. This method reconstructs the iterative process of the Max-Log-MAP algorithm into a multi-level cascaded neural network, combines the Adaptive Prior Enhanced Neural Network APENet to nonlinearly optimize the extrinsic information output by each SISO decoder, and finally realizes high-precision channel decoding. The APENet neural sub-network of this method is a lightweight residual neural network module, which is used to dynamically correct the extrinsic information deviation in traditional linear calculations, improves the modeling ability of the model for complex channel responses, thereby realizing a more accurate estimation of the prior information for the interaction between the two decoders, and finally improving the bit error rate performance of the Turbo decoding model.

[0113] As Figure 1 shown, this method includes the following steps:

[0114] S1: Generate the training data set;

[0115] S2: Construct the APETurbo decoding model based on dual driving of model and data;

[0116] S3: Use the training data set to train the APETurbo decoding model; observe the loss function to ensure that the model converges well, and save the model weight parameters;

[0117] S4: Load the model weight parameters saved in step S3 and complete the model performance test on the test data set.

[0118] This method designs and constructs an intelligent decoding architecture based on the theory of deep learning. This architecture unfolds the M iterations of traditional Turbo decoding into M cascaded neural network levels through deep unfolding. Generally, M is taken as 3, that is, 3 iterations. The neural sub-network APENet is designed in each level. This sub-network model consists of a linear calculation module and a non-linear feature extraction module, with a total of 6 layers: the first layer is the input layer, the middle 4 layers are the hidden layers, and the last layer is the output layer. The sub-network input is the log-likelihood ratio information of the posterior probability L(u k |y), systematic information and prior information L(u k)。The output layer constructs a residual connection structure, linearly calculates the external information of the output as the reference path, and the output value of the non-linear feature extraction module performs residual compensation on the reference path through the adaptive mixing coefficient α, finally realizing the dynamic correction of the prior information deviation, thereby improving the decoding accuracy of the model.

[0119] The following is the simulation implementation process of the method of the present invention:

[0120] As Figure 1 shown, the method specifically includes the following steps:

[0121] S01: Generate a binary information sequence u = (u1, u2,..., u k ) at the sending end, and this sequence is interleaved by an interleaver with a depth of K into Two sequences are respectively encoded for the information sequence through two recursive systematic convolutional (RSC) encoders. The first RSC encoder E1 generates a parity bit sequence according to the information sequence u The second RSC encoder E2 generates a parity bit sequence according to the interleaved information sequence Generate a parity bit sequence Puncture and multiplex the parity sequences according to the code rate requirement to obtain the encoding result. When the code rate is 1 / 3, the encoded codeword consists of N = 3K bits where

[0122] S02: The encoded codeword is transmitted through an additive white Gaussian noise (AWGN) channel after BPSK modulation.

[0123] S03: The received sequence is represented by y = (y1, y2,..., y k ), calculate the reliability information of the transmitted bits in the form of log-likelihood ratios (LLRs) to obtain the channel soft output, representing the probability that the corresponding bit is binary 0 or 1. After derivation, the channel soft output is L c y k , where is a value only related to the signal-to-noise ratio and the channel fading amplitude. After demultiplexing, the system information and the two parity information are respectively represented by and respectively, and the obtained sequence where

[0124] S04: Input the sequence y de obtained in S03 into the Log-MAP decoder for iterative decoding, and obtain the log-likelihood ratio of the posterior probability of the signal output in the last iteration.

[0125] S05: Combine the channel soft output obtained at the receiver, the information bits transmitted by the transmitter, and the log-likelihood ratio of the posterior probability of the information bits obtained by decoding using the Log-MAP decoding algorithm into a set of data as a training data. Collect 60,000 training data to obtain a training data set. Among them, demultiplexing the channel soft output will obtain the system information input to the APETurbo decoder. Check information And check information The information bits transmitted by the transmitter are used to calculate the bit error rate of the training set at the end of each step of training; the log-likelihood ratio of the posterior probability obtained by decoding using the Log-MAP algorithm is used as the training objective for optimizing the APETurbo decoding model.

[0126] S06: Construct an intelligent Turbo channel decoding network model APETurbo decoding model based on dual-driving of model data, as Figure 2 shown. The designed APETurbo decoding model reconstructs the M-iteration process of traditional Turbo channel decoding into an M-layer cascaded neural network architecture through deep unfolding. Generally, M is taken as 3, that is, 3 iterations, and each iteration level is represented by an intelligent decoding unit. As Figure 3 shown, the main structure of each intelligent decoding unit includes: one or more intelligent decoding units with the same structure and an auxiliary hard decision maker. Each intelligent decoding unit includes two identical interleaving units, two identical SISO decoders, a deinterleaving unit corresponding to the interleaving unit, and two APENet networks with the same structure but different weight parameters. In each intelligent decoding unit, the modules of the two SISO decoders and their corresponding APENet neural sub-networks are respectively used to calculate the extrinsic information of the normal layer and the interleaved layer.

[0127] As Figure 4 shown, the APENet neural sub-network designed in the APETurbo decoding model is composed of a linear part calculation module and a prior information non-linear feature extraction module, with a total of 6 layers: the first layer is the input layer, the middle 4 layers are hidden layers, and the last layer is the output layer. The specific structure of each layer of the APENet neural sub-network is as follows:

[0128] K represents the length of the information sequence. The input layer consists of 3K neurons, and the log-likelihood ratio of the posterior probability information L(u k |y), system information and prior information L(u k ).

[0129] The first hidden layer is a linear part calculation layer: Define a triple learnable weight matrix, which acts on the three inputs of the APENet neural sub-network respectively. Then, through bitwise weight sharing of the triple learnable weight matrix, three learnable weights w1, w2, and w3 to be trained in the first layer of the neural sub-network model are obtained. The output of hidden layer 1 is the external information obtained by linear combination calculation after adjustment by learnable parameters. The specific expression is as follows:

[0130]

[0131] Among them, L e (u k ) is the external information obtained by linear calculation after adjustment by learnable parameters, and L(u k |y) is the posterior probability log-likelihood ratio information output by the SISO decoder. is the system information, L(u k ) is the prior information. Define w1, w2, and w3 as the learnable weights acting on the three inputs of the first hidden layer; the output data shape of the first hidden layer is 1×K;

[0132] The second hidden layer is a data flattening layer, and the output data shape of the second hidden layer is K×1;

[0133] The third and fourth hidden layers are non-linear feature extraction layers: Among them, the third hidden layer uses a fully connected layer with 1 input dimension and 64 output dimensions. The weight matrix shape is 64×1, and the output shape is K×64. After the fully connected operation, the PReLU non-linear activation function is used; the fourth hidden layer uses a fully connected layer with 64 input dimensions and 1 output dimension, followed by the Tanh non-linear activation function, and the output is the non-linear feature extraction term of the external information obtained by neural network processing;

[0134] The output layer restores the original data dimension 1×K through data reshaping, constructs a residual connection structure, mixes the linear combination calculation result with the learnable parameters and the non-linear feature extraction term of the external information obtained by neural network processing, uses the output of the linear part calculation layer as the reference path, and uses the non-linear correction term output by the prior information non-linear feature enhancement module as the residual path to achieve gradient stability and feature enhancement through skip connection. The expression is as follows:

[0135] L e_NN (u k )=L e (u k )+α·L e_NNout (u k )

[0136] Among them, α is a learnable mixing coefficient, α∈[0,1], initialized to 0.5, Le (u k ) represents the external information obtained through linear calculation, L e_NNout (u k ) represents the non-linear feature extraction term obtained by enhancing the features of L e (u k ) through neural network, and L e_NN (u k ) represents the corrected external information obtained by combining the linear calculation part and the non-linear feature extraction part in a certain proportion.

[0137] S07: Train the APETurbo decoding model based on model data dual drive. The system information in the training dataset check information check information and the prior information initialized to 0 are input into the APETurbo decoding model for decoding. Assume there are a total of M intelligent decoding units, and the decoding process of the APETurbo decoding model is as follows:

[0138] In the first intelligent decoding unit, the received and are sent to the SISO decoder D1 for decoding to obtain the posterior probability log-likelihood ratio information of each information bit and are sent to the APENet neural sub-network 1 together; the result output by sub-network 1 is the more accurately estimated external information which, after interleaving, becomes the prior information input to the SISO decoder D2 The and the interleaved system information are sent to the SISO decoder 2 for decoding to obtain the posterior probability log-likelihood ratio information of each information bit in the interleaved sequence The interleaved system information and are sent to the APENet neural sub-network 2 together; the result output by sub-network 2 is the more accurately estimated external information which, after de-interleaving, becomes the prior information required for input to the SISO decoder D1 in the next intelligent decoding unit

[0139] In the subsequent m-th intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent to the SISO decoder D1 for decoding to obtain the posterior probability log-likelihood ratio information of each information bit and They are sent into the APENet neural sub-network m together; the result output by sub-network m is the more accurately estimated external information After interleaving, the prior information input to the SISO decoder D2 is obtained The and the interleaved system information are sent into the SISO decoder D2 together for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved system information and are sent into the APENet neural sub-network 2m together; the result output by sub-network 2m is the more accurately estimated external information After de-interleaving, the prior information input to the SISO decoder D1 required by the next intelligent decoding unit is obtained

[0140] In the Mth intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent into the SISO decoder D1 for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and are sent into the APENet neural sub-network M together; the result output by sub-network M is the more accurately estimated external information After interleaving, the prior information input to the SISO decoder D2 is obtained The and the interleaved system information are sent into the SISO decoder D2 together for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence After de-interleaving, the log-likelihood ratio L of the posterior probability of each information bit in the original information sequence is obtained M (u k │y), and the sigmoid function is used to normalize the log-likelihood ratio of the posterior probability of each information bit, and o k is output. For o k a hard decision is made to obtain the final decoded sequence

[0141] In each training cycle, the APETurbo decoding model will perform the above processing on the input data to obtain the log-likelihood ratio of the posterior probability of each bit output by the last intelligent decoding unit of each data. The log-likelihood ratio of the posterior probability of one batch of data is used as the model prediction value; the log-likelihood ratio of the posterior probability of each bit obtained by decoding this batch of data using the Log-MAP algorithm is used as the true value.

[0142] Normalize the predicted values and the true values using the Sigmoid function, calculate the mean squared error loss for all data in a batch, sum them up and take the average; perform regularization constraints on the mixing coefficient α to obtain the regularization constraint term; define a mixed loss function to train and update the parameters of the network APETurbo network model. Observe the loss function to ensure that the model converges well and save the model weight parameters. When the number of training epochs reaches the set number, the training ends, otherwise continue the training for the next epoch.

[0143] Among them, the expression of the mixed loss function required for training the APETurbo decoding model is as follows:

[0144]

[0145] Among them, the first part is the mean squared error (MSE) loss function of a batch of data. σ(·) represents the Sigmoid function, which maps the log-likelihood ratio to the probability space; assume that the training batch size is B, that is, there are B message sequences u (1) , u (2) , …, u (B) , u (i) ∈ {0, 1} K , i = {1, 2, ..., B}, then y (1) , y (2) , …, y (B) is the log-likelihood ratio information of the received information bits, is the predicted value of the log-likelihood ratio of the posterior probability for each bit in each training data; is the true value. The second part is the regularization constraint on the mixing coefficient α. λ is the regularization strength hyperparameter, which is used to constrain the numerical stability of the learnable mixing coefficient α. θ j represents the j-th learnable mixing coefficient α in the entire network.

[0146] The saved weight parameters include: the log-likelihood ratio information L(u k |y) of the posterior probability acting on the output of the SISO decoder, the system information the prior information L(u k ) of the three weights w1, w2, w3; the weights and biases of the fully connected layer; the learnable parameter a in the PReLU non-linear activation function; the learnable mixing coefficient α of the output layer, α ∈ [0, 1].

[0147] S08: As Figure 5As shown in the figure, the APE Turbo decoding model is deployed in the Turbo channel encoding and decoding transmission system described in S01 to S04, and a test data set required to test the performance of the APE Turbo decoding model is generated and the performance of the APE Turbo decoding model is tested. First, an information bit sequence is generated by the transmitter of the Turbo channel encoding and decoding transmission system, and the information sequence is Turbo encoded using a parallel concatenated convolutional code structure to generate an encoded sequence. The encoded sequence is modulated and transmitted through an AWGN channel, adding Gaussian white noise with a specified signal-to-noise ratio, and sent to the receiver.

[0148] The receiver receives the noisy information sequence, demodulates it, and calculates the channel soft output. Demultiplexing the channel soft output will obtain the system information input to the APE Turbo decoder parity information and the parity information together with the information bits sent by the transmitter form a set of data, which is used as test data in a specific signal-to-noise ratio environment. 5,000,000 training data are collected to obtain a training data set in a specific signal-to-noise ratio environment.

[0149] Collect training data sets in multiple signal-to-noise ratio environments. The system information parity information and the parity information in the training data set collected in a specific signal-to-noise ratio environment are input into the APE Turbo decoding model. The prior information input to the model is initialized to 0. After being processed by the APE Turbo decoding model, a decoded sequence is obtained. According to the information bits sent by the transmitter, the bit error rate of the test data set of the information bits sent by the transmitter in each signal-to-noise ratio environment is calculated to obtain a bit error rate performance curve, as Figure 6 shown. The comparison algorithms are the Max-Log-MAP algorithm (denoted as "MaxLogMAP"), the Log-MAP algorithm (denoted as "MAP"), the model-driven TinyTurbo intelligent channel decoding algorithm (denoted as "TinyTurbo"), and the model-data dual-driven APE Turbo intelligent channel decoding algorithm (denoted as "APE Turbo"). When using a code length of 40 and a code rate with three iterations, the model-data dual-driven APE Turbo intelligent channel decoding algorithm proposed by the present invention has better performance than the corresponding Turbo traditional channel decoding algorithms, the Max-Log-MAP algorithm and the Log-MAP algorithm; and the bit error rate performance of the APE Turbo decoder proposed by the present invention is better than the performance of the existing model-driven TinyTurbo intelligent channel decoding algorithm.

[0150] The neural network model of the present invention adopts the method of offline training and online testing, transfers the computational amount and overhead to the offline training process, and greatly reduces the overhead in the actual application of the algorithm. Both theoretical analysis and simulation results prove that the present invention has excellent performance beyond the traditional Turbo channel decoding algorithm and the existing model-driven intelligent Turbo decoding algorithm. When the signal-to-noise ratio is 3 dB, the bit error rate performance can be improved by about half an order of magnitude.

[0151] Compared with the traditional Turbo decoding algorithm, the present invention significantly reduces the bit error rate. Compared with the existing model-driven Turbo decoding algorithm, the present invention retains the traditional linear calculation and adds non-linear feature extraction, while improving the non-linear feature modeling ability and maintaining the interpretability of the algorithm, further improving the decoding accuracy of the model.

[0152] In summary, the present invention aims at the problems of insufficient bit error rate performance of the traditional Turbo channel decoding method, as well as the problems that the performance of the existing model-driven Turbo channel decoding method still has room for improvement, the interpretability of the data-driven channel decoding method is insufficient, and the model complexity is relatively high. By designing and building an intelligent Turbo decoding model APETurbo based on model-data dual drive, the decoding result accuracy is further improved.

Claims

1. An intelligent Turbo channel decoding method based on dual drive of model and data, characterized in that It includes the following steps: S1: Generate a training data set; S2: Construct an APETurbo decoding model driven by both model and data; S3: Use the training data set to train the APETurbo decoding model and save the model weight parameters; S4: Load the model weight parameters saved in step S3, and deploy the APETurbo decoding model at the receiving end of the communication system for channel decoding.

2. The intelligent Turbo channel decoding method based on dual drive of model and data according to claim 1, wherein The specific method of step S1 is as follows: S11: Build a Turbo channel encoding and decoding transmission system in the Pycharm simulation platform. The Turbo channel encoding and decoding transmission system includes a sending end, an AWGN channel, and a receiving end; S12: Generate an information bit sequence through the sending end of the Turbo channel encoding and decoding transmission system; then, perform Turbo encoding on the information sequence using a parallel concatenated convolutional code structure to generate a coded sequence; S13: Modulate the coded sequence, transmit the information bits through the AWGN channel, add Gaussian white noise to the channel, and send it to the receiving end; S14: The receiving end receives the information sequence with added Gaussian white noise, demodulates the information sequence, calculates the channel soft output, decodes it using the Log-MAP algorithm, and obtains the log-likelihood ratio of the posterior probability of the signal output in the last iteration; S15: Combine the channel soft output obtained at the receiving end, the information bits sent by the sending end, and the log-likelihood ratio of the posterior probability obtained in step S14 into a set of data as a training data; S16: Collect multiple training data to obtain a training data set.

3. The intelligent Turbo channel decoding method based on dual drive of model and data according to claim 1, characterized in that The APETurbo decoding model includes one or more intelligent decoding units with the same structure and an auxiliary hard decision maker. Each intelligent decoding unit includes two identical interleaving units, two identical SISO decoders, a deinterleaving unit corresponding to the interleaving unit, and two APENet networks with the same structure but different weight parameters; The APENet network has 6 layers: the first layer is the input layer, the middle 4 layers are the hidden layers, and the last layer is the output layer; among them: The input layer consists of 3K neurons, representing the log-likelihood ratio information L(u k |y) of K bits, the systematic information and the prior information L(u k ); The first hidden layer is a linear part calculation layer, and the expression is as follows: Among them, L e (u k ) is the external information obtained by linearly calculating through learnable parameter adjustment. L(u k |y) is the log-likelihood ratio information of the posterior probability output by the SISO decoder, is the system information. L(u k ) is the prior information. Define w1, w2, w3 as the learnable weights acting on the three inputs of the first hidden layer; the output data shape of the first hidden layer is 1×K; The second hidden layer is a data flattening layer, and the data shape output by the second hidden layer is K×1; The third and fourth hidden layers are non-linear feature extraction layers: among them, the third hidden layer uses a fully connected layer with an input of 1 dimension and an output of 64 dimensions, the weight matrix shape is 64×1, the output shape is K×64, and after the fully connected operation, the PReLU non-linear activation function is used; the fourth hidden layer uses a fully connected layer with an input dimension of 64 and an output of 1 dimension, followed by the Tanh non-linear activation function, and the output is the non-linear feature extraction term of the external information obtained through neural network processing; The output layer restores the original data dimension of 1×K through data reshaping, constructs a residual connection structure, and mixes the calculation result of the linear combination with the learnable parameters and the non-linear feature extraction term of the external information obtained through neural network processing. The expression is as follows: L e_NN (u k ) = L e (u k ) + α·L e_NNout (u k ) where α is a learnable mixing coefficient, α ∈ [0, 1], L e (u k ) represents the external information obtained through linear calculation, L e_NNout (u k ) represents the non-linear feature extraction term obtained by enhancing the features of L e (u k ) using a neural network, and L e_NN (u k ) represents the corrected external information obtained by combining the linear calculation part and the non-linear feature extraction part in a certain proportion.

4. A method for intelligent Turbo channel decoding based on dual-driving of model and data according to claim 3, characterized in that The input of the APE Turbo decoding model is system information the first parity information the second parity information and the prior information Among them, the system information, the first parity information, and the second parity information are respectively input into each intelligent decoding unit. The initialized to 0; Suppose there are M intelligent decoding units in total, and each intelligent decoding unit includes a first SISO decoder, a second SISO decoder, a first APENet network, and a second APENet network; In the first intelligent decoding unit, the received and are sent to the first SISO decoder for decoding to obtain the posterior probability log-likelihood ratio information of each information bit and which are sent to the first APENet network together; the result output by the first APENet network is the more accurately estimated extrinsic information which, after interleaving, becomes the prior information input to the second SISO decoder The and the interleaved systematic information are sent to the second SISO decoder for decoding to obtain the posterior probability log-likelihood ratio information of each information bit in the interleaved sequence The interleaved systematic information and are sent to the second APENet network together; the result output by the second APENet network is the more accurately estimated extrinsic information which, after deinterleaving, becomes the prior information required to be input to the first SISO decoder of the next intelligent decoding unit In the subsequent m-th intelligent decoding unit, where m = 2, 3,..., M - 1, the received and the prior information output by the previous intelligent decoding unit are fed into the first SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit and which are then fed into the first APENet network together; the result output by the first APENet network is the more precisely estimated extrinsic information which, after interleaving, becomes the prior information input to the second SISO decoder The and the interleaved systematic information are fed into the second SISO decoder for decoding to obtain the log-likelihood ratio information of the posterior probability of each information bit in the interleaved sequence The interleaved systematic information and are fed into the second APENet network together; the result output by the second APENet network is the more precisely estimated extrinsic information which, after deinterleaving, becomes the prior information required for input to the first SISO decoder of the next intelligent decoding unit In the Mth intelligent decoding unit, the received and the prior information output by the previous intelligent decoding unit are sent to the first SISO decoder for decoding to obtain the posterior probability log-likelihood ratio information of each information bit and are sent to the first APENet network together; the result output by the first APENet network is the more accurately estimated extrinsic information After interleaving, the obtained prior information is input to the second SISO decoder The and the interleaved system information are sent to the second SISO decoder for decoding to obtain the posterior probability log-likelihood ratio information of each information bit in the interleaved sequence The interleaved system information and are sent to the second APENet network together; the result output by the second APENet network is the more accurately estimated extrinsic information After deinterleaving, it is added to the system information to obtain the posterior probability log-likelihood ratio L of each information bit in the original information sequence M (u k │y), and the sigmoid function is used to normalize the posterior probability log-likelihood ratio of each information bit, and the output is o k For o k a hard decision is made to obtain the final decoded sequence 5. The intelligent Turbo channel decoding method based on dual drive of model and data according to claim 3, characterized in that The weight parameters of the APENet network include: the log-likelihood ratio information of the posterior probability acting on the output of the SISO decoder L(u k |y), the systematic information prior information L(u k ), and three weights w1, w2, w3; the weights and biases of the fully connected layer; the learnable parameter a in the PReLU non-linear activation function; the learnable mixing coefficient α of the output layer, where α ∈ [0, 1].

6. The intelligent Turbo channel decoding method based on dual drive of model and data according to claim 2, wherein The training process of the APETurbo decoding model is as follows: (1) Demultiplex the channel soft output in the training dataset to obtain system information Parity check information And parity check information Will And the prior information initialized to 0 Input into the APETurbo decoding model for decoding; (2) In each training cycle, the APETurbo decoding model processes the input data and outputs the log-likelihood ratio of the posterior probability of each bit through the last intelligent decoding unit; the log-likelihood ratio of the posterior probability of a batch of data is used as the model prediction value, and the log-likelihood ratio of the posterior probability of each bit obtained by the Log-MAP algorithm decoding of this batch of data in the training dataset is used as the model true value; (3) Normalize the prediction value and the true value using the Sigmoid function, calculate the mean square error loss of all data in a batch, sum and take the average; perform regularization constraint on the mixing coefficient α to obtain the regularization constraint term; define a mixed loss function to train and update the parameters of the APETurbo network model; the expression of the mixed loss function is as follows: Among them, is the mean squared error loss function of a batch of data, and σ(·) represents the Sigmoid function, which maps the log-likelihood ratio to the probability space. Assuming the training batch size is B, that is, there are B message sequences u of length K (1) , u (2) , …, u (B) , u (i) ∈ {0, 1} K , i = 1, 2, ..., B, then y (1) , y (2) , …, y (B) is the log-likelihood ratio information of the received information bits; is the predicted value of the posterior probability log-likelihood ratio for each bit in each training data; is the true value; is the regularization constraint on the mixing coefficient α, and λ is the regularization strength hyperparameter, which is used to constrain the numerical stability of the learnable mixing coefficient α, and θ j represents the j-th learnable mixing coefficient α in the entire network; (4) Observe the loss function. When the model converges or reaches the maximum number of training times, the training is completed and the weight parameters of the model are saved.

7. A method for intelligent Turbo channel decoding based on dual drive of model and data according to claim 1, characterized in that The specific method of step S4 is as follows: S41: Load the saved model parameters and deploy the APETurbo decoding model at the receiving end of the communication system; S42: The receiving end receives the noisy information sequence transmitted from the sending end, demodulates the noisy information sequence, and calculates the channel soft output; S43: Demultiplex the channel soft output to obtain the system information, the first parity check information, and the second parity check information, and input these information together with the prior information initialized to 0 into the APETurbo decoding model for channel decoding to obtain the decoded information sequence.

Images