A Deep Learning-Based Method for Cleavage Convolutional Code Rate Recognition
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-29
- Publication Date
- 2026-08-14
AI Technical Summary
[0004]针对现有的分类树识别方法中人工提取复杂度高，深度学习算法框架没有考虑删余卷积码相关性特征的问题，本发明的主要目的是提出一种基于深度学习的删余卷积码码率识别方法，通过多尺度卷积网络结合删余卷积码的线性相关性，提高信道编码的识别率、降低运算复杂度
[0021]1、本发明公开的一种基于深度学习的删余卷积码码率识别方法，建立基于一维多尺度卷积网络的识别模型，直接将一维的比特数据输入网络，输出分类识别结果，避免人工提取特征的复杂性。
Smart Images
Figure CN116488663B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a deep learning-based method for identifying the bit rate of pruned convolutional codes, belonging to the field of channel coding parameter identification technology. Background Technology
[0002] Channel coding improves the reliability of communication system transmission, and convolutional codes are an important channel coding method employing a tree code structure. Due to their efficient and fast Viterbi decoding algorithm, they are widely used in many systems such as satellite communication and digital cellular mobile communication. Pruned convolutional codes are constructed using different pruning matrices to improve the efficiency of convolutional codes. Typical pruned convolutional codes are 2 / 3, 3 / 4, 4 / 5, 5 / 6, 6 / 7, and 7 / 8 pruning codes constructed from convolutional codes with a 1 / 2 code rate.
[0003] In certain applications, such as cognitive radio and radio spectrum detection, it is necessary to obtain channel coding parameters through signal analysis without prior information. Essentially, identifying channel coding parameters is a pattern recognition problem, and deep learning networks are an important tool in this field. Currently, the channel coding used in communication systems, such as convolutional codes and LDPC codes, are linear correlation codes. Compared to traditional feature-based classification tree recognition methods, deep learning frameworks can automatically extract correlation features from raw data, avoiding the complexity of manual feature extraction. While there has been extensive research on deep learning methods in the field of channel coding recognition, research on the identification of convolutional code parameters with different code rates generated by pruning remains relatively rare. Summary of the Invention
[0004] To address the issues of high complexity in manual extraction and the lack of consideration for correlation features in pruned convolutional codes in existing classification tree recognition methods, the main objective of this invention is to propose a deep learning-based method for identifying the code rate of pruned convolutional codes. This method improves the recognition rate of channel coding and reduces computational complexity by combining multi-scale convolutional networks with the linear correlation of pruned convolutional codes.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] This invention discloses a deep learning-based method for identifying the bitrate of pruned convolutional codes, comprising the following steps:
[0007] Step 1: Determine the candidate set: Based on the known convolutional code, the pruned convolutional code is obtained, and the code rate candidate set is θ = {C1, ..., C...} KLet K be the number of code rates, and the candidate code rates include 1 / 2, 2 / 3, ..., L-1 / L, where L is a positive integer. Construct pruning matrices corresponding to different code rates, where each bit in the pruning matrix retains at least one bit of information. The deep learning model identifies the N×1 information bit sequence of the input model, where N is the number of bits in the signal, i.e., its length.
[0008] Step 2: Generate Training Set: Based on the various bitrates in the bitrate candidate set from Step 1, generate m training data sets r = (r1, ..., r2) for each bitrate according to different pruning matrices. N The starting point is known as the starting point of the codeword. Each training data point is labeled accordingly and used as the training set to be input into the model for training.
[0009] Step 3: Generate validation and test sets: Based on the various bitrates in the bitrate candidate set from Step 1, generate m' training data sets r = (r1, ..., r2) for each bitrate according to different pruning matrices. N The starting point is known as the starting point of the codeword. Each training data point is labeled accordingly, serving as the validation set and test set.
[0010] Step 4: Building the Network Model: A network model is built based on a one-dimensional convolutional network. Residual blocks and ensemble learning are introduced into the model. A multi-scale feature extractor is added between the input layer and the residual blocks, and a global average pooling layer is added before the output layer. The output layer uses the Softmax activation function to output the confidence probability of the signal bitstream corresponding to each category. Finally, multiple models with different initializations are integrated, and the recognition accuracy is improved through voting.
[0011] 4.1: The multi-scale feature extractor consists of four convolutional kernels of different sizes, selected as 2, 3, 4, and 5 respectively. These kernels extract the lowest-level features of the data, and the extracted features are then superimposed to preserve information at different scales. After low-level feature extraction by the multi-scale feature extractor, the data is input into the residual block for high-dimensional feature extraction.
[0012] 4.2: Residual blocks are formed by connecting two dilated convolutions. The input of the residual block is connected to the output in a skip manner, thus enabling the direct transmission of information to deeper layers of the model. The activation function of the two convolutional layers is the Rectified Linear Unit (ReLU). The convolutional layers increase the receptive field of the convolutional kernel through dilated convolutions; the larger the dilation coefficient, the longer the sequence the convolutional kernel can cover. The receptive field of a convolution with a dilation coefficient of 1, i.e., a regular convolution, is:
[0013]
[0014] Among them l k and f kTo determine the size of the receptive field and the convolution kernel k, s i This represents the step size of the convolutional kernel in the i-th convolutional layer. The dilation factor d refers to inserting d-1 zeros into the convolutional kernel; therefore, the kernel size of dilated convolution is:
[0015]
[0016] in This indicates the size of the convolution kernel after dilation.
[0017] 4.3: Use the above models with different initializations for ensemble learning, identify the same signal to be classified and output the results, and count the voting results to improve the recognition accuracy of the model in the form of voting. The classification that most models consider correct is considered the final classification result.
[0018] Step 5: Train the network model: Feed the training dataset into the network for training, and use the validation set to adjust the model's hyperparameters to obtain a well-trained network model. After training, this deep learning model can automatically extract features from the original data using a multi-scale convolutional network and assign signal bitstreams to different bitrates with confidence probabilities. Use the test set to evaluate the model's learning performance.
[0019] Step Six: Bitrate Identification: The received sequence of information to be identified, x = (x1, ..., x...), is... N The input is given to the trained network for recognition, and the recognition result C is obtained from the network. i ∈θ. Improve the recognition rate of channel coding and reduce computational complexity.
[0020] Beneficial effects:
[0021] 1. The present invention discloses a deep learning-based method for identifying the bit rate of pruned convolutional codes. It establishes a recognition model based on a one-dimensional multi-scale convolutional network, directly inputs one-dimensional bit data into the network, and outputs classification and recognition results, thus avoiding the complexity of manual feature extraction.
[0022] 2. The present invention discloses a deep learning-based method for identifying the bitrate of pruned convolutional codes. It establishes an identification model based on a one-dimensional multi-scale convolutional network. Based on the linear correlation of pruned convolutional codes, it adopts a multi-scale convolutional network framework that can better identify the features of the pruned matrix. Compared with other existing deep learning frameworks, it can achieve a higher recognition rate and better anti-interference ability.
[0023] 3. The present invention discloses a deep learning-based method for identifying the bitrate of pruned convolutional codes. By employing an ensemble learning approach, the method simultaneously inputs data into multiple models with different initializations for identification and then votes to obtain the results. This effectively improves the robustness of the models, making them more resistant to interference and more stable in the face of different input data. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating a deep learning framework recognition method for pruning convolutional codes disclosed in this invention.
[0025] Figure 2 This is a diagram of the deep learning model used in this embodiment;
[0026] Figure 3 This is a schematic diagram of the residual block in this embodiment;
[0027] Figure 4 This is a schematic diagram of the dilated convolution structure in this embodiment;
[0028] Figure 5 This is a schematic diagram of the multi-scale feature extractor in this embodiment;
[0029] Figure 6 This is a schematic diagram comparing the recognition rates of different methods in this embodiment. Detailed Implementation
[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical problems solved by the present invention and its beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not constitute any limitation thereof.
[0031] like Figure 1 As shown, this embodiment discloses a deep learning-based method for identifying the code rate of pruned convolutional codes, which is used to solve cognitive radio and radio spectrum detection scenarios. This method requires obtaining channel coding parameters through received signal analysis without prior information, and includes the following steps:
[0032] Step 1: Determine the candidate set: Cleave the known convolutional code (2,1,7) to obtain the pruned convolutional code. The generator polynomial of the (2,1,7) convolutional code is as follows:
[0033] g1 = 1011011
[0034] g2 = 1111001
[0035] The candidate set of bitrates is θ = {C1, ..., C...} K Let K be the number of code rates. In this example, K = 4, and the candidate code rates include 1 / 2, 2 / 3, 3 / 4, and 4 / 5. Determine the pruning matrix corresponding to different code rates. The pruning matrix is shown in the table below.
[0036]
[0037]
[0038] The deep learning model will identify the obtained information bit sequence N×1, where N is the number of bits in the signal, i.e., its length. This example uses an information bit sequence of 1500x1.
[0039] Step 2: Generate the training set: Based on the various bitrates in the bitrate candidate set from Step 1, generate 15,000 training data points r = (r1, ..., r2) for each bitrate according to different pruning matrices. N The starting point is known as the starting point of the codeword. Each training data point is labeled accordingly and used as the training set to be input into the model for training.
[0040] Step 3: Determine the validation and test sets: Based on the various bitrates in the bitrate candidate set from Step 1, generate 1000 training data points r = (r1, ..., r2) for each bitrate according to different pruning matrices. N The starting point is known as the starting point of the codeword. Each training data point is labeled accordingly, serving as the validation set and test set.
[0041] Step 4: Building the Network Model: A network model is built based on a one-dimensional Convolutional Neural Network (CNN), introducing the concept of residual blocks. A multi-scale feature extractor is added between the input layer and the residual blocks, and a Global Average Pooling (GAP) layer is added before the output layer. Furthermore, this model incorporates ensemble learning, integrating multiple models with different initializations to improve the model's recognition accuracy through voting. The specific structure of the model is as follows... Figure 2 As shown.
[0042] In this example, a residual block is defined as a concatenation of two dilated convolutions. Furthermore, the input to the residual block is connected to the output in a skip manner, thus passing information to deeper layers of the model. The two convolutional layers in this example use the Rectified Linear Unit (ReLU) activation function. The residual block design is as follows: Figure 3 As shown.
[0043] Convolutional layers increase the receptive field of the convolution kernel through dilated convolution; the larger the dilation coefficient, the longer the sequence the kernel can cover. The receptive field of a convolution with a dilation coefficient of 1, i.e., a regular convolution, is:
[0044]
[0045] Among them l k and f k To determine the size of the receptive field and the convolution kernel k, s iThis represents the step size of the convolution kernel in the i-th convolutional layer. The dilation factor d is the number of zeros inserted into the convolution kernel (d-1), therefore the kernel size of dilated convolution is:
[0046]
[0047] in This indicates the kernel size after dilated convolution. It's clear that dilated convolution significantly improves the receptive field. Figure 4 As shown.
[0048] Design of multi-scale feature extractors, such as Figure 5 As shown, the dataset consists of four convolutional kernels of different sizes, with lengths of 2, 3, 4, and 5. These kernels are used to extract the lowest-level features of the data, and the extracted features are then stacked together to preserve information at different scales. The low-level features are then fed into the residual block for high-dimensional feature extraction.
[0049] The ensemble learning involved in this invention is as follows: the above models with different initializations are integrated to identify the same signal to be classified, the voting results are statistically analyzed, and the classification that most models consider correct is considered the final classification result.
[0050] Step 5: Train the network model: Feed the training dataset into the network for training, and use the validation set to adjust the model's hyperparameters to obtain a trained network model. This deep learning model can automatically extract features from the original data and assign the signal bitstream to different bitrates with confidence probabilities.
[0051] Step Six: Bitrate Identification: The received sequence of information to be identified, i.e., the test set x = (x1, ..., x...), is... N The input is given to the trained network for recognition, and the recognition result C is obtained from the network. i ∈θ.
[0052] Using the method of this invention, the recognition rate of pruned convolutional codes is compared as follows: Figure 6 As shown, the bit error rate ranges from 0 to 5*10^-2. It can be seen that the network model using this invention shows a certain improvement in recognition rate under different signal-to-noise ratios compared to other existing deep learning models, such as ResNet, DNN, RNN, LSTM, CNN, DRN, and Bi-LSTM, and has stronger anti-interference capabilities. In summary, this embodiment mainly proposes a deep learning framework recognition method for pruned convolutional codes, which can be used in various scenarios and achieves good recognition performance.
[0053] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A deep learning-based method for identifying the bitrate of pruned convolutional codes, characterized in that: Includes the following steps, Step 1: Determine the candidate set: Based on the known convolutional codes, perform pruning to obtain the pruned convolutional codes. The code rate candidate set is as follows: K is the number of code rates, and the candidate code rates include 1 / 2, 2 / 3, ..., L-1 / L, where L is a positive integer; construct the pruning matrix corresponding to different code rates, where each bit in the pruning matrix has at least one bit of information retained; The deep learning model's information bit sequence from the input model For identification, N is the number of bits in the signal, i.e., the length; Step 2: Generate Training Set: Based on the various bitrates in the bitrate candidate set from Step 1, generate m training data sets for each bitrate according to different pruning matrices. The starting point is known as the starting point of the codeword; Each training data point is labeled accordingly and used as a training set to be input into the model for training. Step 3: Generate validation and test sets: Based on the various bitrates in the bitrate candidate set from Step 1, generate m' training data sets for each bitrate according to different pruning matrices. The starting point is known as the starting point of the codeword; Each training data point is labeled accordingly, serving as both a validation set and a test set; Step 4: Build the network model: Build a network model based on a one-dimensional convolutional network, and introduce residual blocks and ensemble learning into the model. Add a multi-scale feature extractor between the input layer and the residual blocks, and add a global average pooling layer before the output layer. The output layer uses the Softmax activation function to output the confidence probability of each category corresponding to the signal bit stream. Integrate multiple models with different initializations and improve the recognition accuracy of the model by voting. The implementation method for step four is as follows: 4.1: The multi-scale feature extractor consists of four convolutional kernels of different sizes, with kernel sizes of 2, 3, 4, and 5 selected respectively. These kernels extract the lowest-level features of the data, and the extracted features are then superimposed to preserve information at different scales. After the low-level features are extracted by the multi-scale feature extractor, they are then input into the residual block for high-dimensional feature extraction. 4.2: The residual block is composed of two dilated convolutions connected together. The input of the residual block will jump to the output, thus enabling the direct transmission of information to deeper layers of the model; the activation function of the two convolutional layers is a linear rectified function; the convolutional layers increase the receptive field of the convolutional kernel by dilating the convolutional kernel. The larger the dilation coefficient, the longer the convolutional kernel can cover the sequence; the receptive field of a convolution with a dilation coefficient of 1, i.e., a normal convolution, is: in and To determine the size of the receptive field and the convolution kernel k, This represents the step size of the convolution kernel in the i-th convolutional layer; the dilation coefficient d refers to inserting d-1 zeros into the convolution kernel, therefore the kernel size of dilated convolution is: in This indicates the size of the convolution kernel after dilation. 4.3: Use the above models with different initializations for ensemble learning, identify the same signal to be classified and output the results, and statistically analyze the voting results to improve the recognition accuracy of the model in the form of voting; Step 5: Train the network model: Feed the training dataset into the network for training, and use the validation set to adjust the model's hyperparameters to obtain a trained network model; After training, the deep learning model can use a multi-scale convolutional network to extract features from the original data and assign signal bitstreams to different bitrates with confidence probabilities; Use the test set to evaluate the model's learning performance. Step Six: Bitrate Identification: The received sequence of information to be identified... Input the trained network from step five into the network for recognition, and obtain the recognition results from the network. Improve the recognition rate of channel coding and reduce computational complexity.