CN109525253B - Convolutional code decoding method based on deep learning and integration method - Google Patents

Abstract

The invention provides a convolutional code decoding method based on deep learning and integration method, the method sets the number of weak classifiers and weak classifiers; Decoding, and setting the depth of the deep neural network; finally, the ensemble method is used to vote the decoding result of the weak classifier to obtain the decoding output; the deep neural network is a fully connected neural network, a convolutional neural network, GAN or LSTM . The method of the invention adopts the deep learning algorithm and the integration method to decode the convolutional code, and restores the transmitted information bit sequence from the noisy soft information sequence.

Description

Convolutional code decoding method based on deep learning and ensemble methods

technical field

The present invention relates to the technical field of electronic communication, and more particularly, to a convolutional code decoding method based on deep learning and integration methods.

Background technique

In order to improve the reliability of signal transmission in the channel, various error correction code technologies are widely used in digital communication. Convolutional code (convolutional code) is a widely used coding method with good performance. In transmission systems, especially satellite communication systems, Viterbi is a decoding method for convolutional codes.

Convolutional code was proposed by Elias in 1955. Convolutional code is a kind of cyclic code. The difference between cyclic code and block code is that during the encoding and decoding process of block code, the n-k checksums of this group are The element is only related to the k information elements of this group, and has nothing to do with other groups. In convolutional code encoding and decoding, the n-k check elements of this group are not only related to the k information elements of this group, but also related to the information groups input to the encoder at previous times. Because the correlation between groups is fully utilized in the encoding process of convolutional codes, and k and n are also small, therefore, under the same code rate and equipment complexity as block codes, no matter theoretically It has been proved that the performance of convolutional codes is at least no worse than that of block codes.

For the existing Viterbi decoding methods of convolutional codes, Viterbi decoding still has room for improvement in the balance between decoding efficiency and decoding performance. When the decoding window of Viterbi is fixed, Viterbi decoding passes Calculating the Hamming distance to obtain the optimal path greatly reduces the decoding efficiency.

SUMMARY OF THE INVENTION

The object of the present invention is to overcome the shortcomings and deficiencies in the prior art, and to provide a convolutional code decoding method based on deep learning and integration methods, which uses deep learning algorithms and integration methods to perform convolutional code decoding. Decoding and recovering the transmitted information bit sequence from the noisy soft information sequence.

In order to achieve the above object, the present invention is achieved by the following technical solutions: a convolutional code decoding method based on deep learning and an integrated method, characterized in that: setting the number of weak classifiers and weak classifiers; The classifier uses a deep neural network or a perceptron to decode the convolutional code, and sets the depth of the deep neural network; finally, an ensemble method is used to vote on the decoding result of the weak classifier to obtain a decoding output; the deep neural network Be fully connected neural network, convolutional neural network, GAN or LSTM.

The weak classifier uses the deep neural network to decode the convolutional code, and sets the depth of the deep neural network; finally, the ensemble method is used to vote on the decoding result of the weak classifier to obtain the decoding output: in the weak classifier , by establishing a deep neural network model and dividing the semi-infinite convolutional code sequence into a training set that conforms to the deep neural network structure. After training the deep neural network model, the divided noisy convolutional codes are processed in different dimensions. Decoding; finally, the ensemble vote is used to convert it into the decoding output of all codewords.

Include the following steps:

The first step is to determine the model parameters of the deep neural network in the weak classifier, and establish the deep neural network model;

The second step is to establish a data sample set for convolutional code decoding;

The third step is to use the data sample set of the second step and use the softmax classification method and the batch gradient descent method to train the deep neural network model;

The fourth step is to input the convolutional code to be decoded into the deep neural network model obtained in the third step for decoding; in the decoding process, several weak classifiers are obtained to decode the volume to be decoded in different dimensions. The information bits corresponding to the information code field of the product code are classified and decoded for output, and the decoded output of the information bits is voted by an integrated method to generate a strong classifier to obtain the final decoding and complete the decoding of the convolutional code.

In the first step, determining the model parameters of the deep neural network and establishing the deep neural network model means: for any (n ₀ , k ₀ , m) convolutional code, setting the output layer of the deep neural network model The dimension is n, the dimension of the input layer is n ₀ ×n/k ₀ ; the activation function of the hidden layer is set as f(x)=relu(x); the depth is established according to the dimension of the output layer, the dimension of the input layer and the activation function of the hidden layer Neural network model.

The described data sample set for establishing convolutional code decoding refers to:

First, an information sequence of length L is randomly generated, encoded by (n ₀ , k ₀ , m) convolutional code, and a noisy convolution of length n ₀ ×L/k ₀ is obtained by adding noise to Gaussian white noise code information sequence;

Second, add 00 as the status bit before the noisy convolutional code information sequence, and divide the noisy convolutional code information sequence according to the input dimension of the deep neural network model in the first step to form a size corresponding to the deep neural network model. The information code field of the noisy convolutional code; wherein, the start state of any (n ₀ , k ₀ , m) convolutional code is 00;

Finally, sample construction is performed on the information code field of the noisy convolutional code, and a data sample set conforming to the deep neural network model is generated in batches.

The described sample construction of the information code field of the noisy convolutional code, and batch generation of data sample sets conforming to the deep neural network model refers to:

(1) In the information code field of the noisy convolutional code, the first k ₀ ×m bits are the status bits of the original codeword, and the last n ₀ ×n/k ₀ bits are the information code field of the noisy convolutional code, as the first training samples;

(2) Set the size of the sample acquisition information bit window to N. When taking the second training sample, the sample window slides one bit backward in the sequence direction in the information code field of the noisy convolutional code, and the status bit of the previous code field is set. The second 0 and the first bit of the code field are used as the status bit, plus the codeword bits added to the sample window after sliding, as the second training sample;

(3) By analogy, a data sample set conforming to the deep neural network model is generated in batches according to the entire segment of the noisy convolutional code information code field and the corresponding information bits.

In the third step, during the training of the deep neural network model, two processes of feedforward calculation and backward propagation are used to update the weights to obtain the optimal weights, so that the model has the ability to classify.

In the fourth step, the convolutional code that needs to be decoded is input into the deep neural network model obtained in the third step for decoding; in the decoding process, several weak classifiers are obtained in different dimensions. The information bits corresponding to the information code field of the decoded convolutional code are classified and decoded for output, and an integrated method is used to vote on the decoded output of the information bits to generate a strong classifier to obtain the final decoding and complete the convolutional code decoding. Code means:

(1) Encode the convolutional code to be decoded and add noise to obtain a noisy convolutional code information sequence, and input the deep neural network model after adding zero information bits to the information bits of the noisy convolutional code information sequence at the end;

(2) Set the initial state bit to 00, and decode the first information bit;

(3) Update the state of the noisy convolutional code information sequence, slide one bit to the back of the noisy convolutional code information sequence, and get a decoding output for each one bit; slide to the back of the noisy convolutional code information sequence When the number of times reaches the output layer dimension size n of the deep neural network model, n weak classifiers including the information bit are obtained, and these n weak classifiers classify the information bit in different dimensions and have n Sub-decoding output; use the integrated method to vote on this information bit to generate a strong classifier: that is, perform statistics on the n-time decoding output, and vote to obtain a large number of decoding outputs as the decoding results of the information bit;

(4) Repeat this step of decoding for the subsequent noisy convolutional code information sequence to complete the convolutional code decoding.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The convolutional code decoding method based on the deep learning and integration method of the present invention adopts the deep learning algorithm and the integration method to decode the convolutional code, and restore the transmitted information bit sequence from the noisy soft information sequence.

2. The present invention obtains a plurality of weak classifiers through a deep neural network model, and integrates them into a strong classifier, which greatly improves the decoding performance of the deep neural network model.

Description of drawings

1 is a flowchart of a convolutional code decoding method based on a deep learning and integration method according to an embodiment of the present invention;

2 is a structural diagram of a deep neural network model according to Embodiment 1 of the present invention;

Fig. 3 is the decoding process diagram of the weak classifier of the method of Embodiment 1 of the present invention;

FIG. 4 is a performance comparison diagram of decoding and Viterbi decoding of a method according to Embodiment 1 of the present invention.

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Example 1

The present invention takes (2, 1, 2) convolutional code as an embodiment, and provides a detailed description of a convolutional code decoding method based on deep learning and integration methods provided by the present invention, wherein the encoding method of the convolutional code is as follows , the initial state of each segment of the convolutional code can be represented by two input bits before the segment, and the initial state at the front end of the code can be represented by 00. G(D)=[1+D+D2,1+D2].

As shown in Figures 1 to 4, the convolutional code decoding method based on the deep learning and ensemble method of the present invention is as follows: set the number of weak classifiers and weak classifiers; The product code is decoded, and the depth of the deep neural network is set; finally, the ensemble method is used to vote the decoding result of the weak classifier to obtain the decoding output; the deep neural network is a fully connected neural network, a convolutional neural network, GAN or LSTM.

Specifically, by establishing a deep neural network model and dividing the semi-infinite convolutional code sequence into a training set that conforms to the deep neural network structure, after training the deep neural network model, the divided noisy convolutional codes are processed in different dimensions. , and finally use the ensemble method to vote to convert it into the decoding output of all codewords.

Include the following steps:

The first step is to determine the model parameters of the deep neural network in the weak classifier, and establish the deep neural network model;

The second step is to establish a data sample set for convolutional code decoding;

The third step is to use the data sample set of the second step and use the softmax classification method and the batch gradient descent method to train the deep neural network model;

The fourth step is to input the convolutional code to be decoded into the deep neural network model obtained in the third step for decoding; in the decoding process, several weak classifiers are obtained to decode the volume to be decoded in different dimensions. The information bits corresponding to the information code field of the product code are classified, and the information bits are voted by an integrated method to generate a strong classifier, which is decoded by the strong classifier to complete the decoding of the convolutional code.

Specific steps are as follows:

(1) First determine the model parameters of the deep neural network, and establish the deep neural network model. The dimension of the output layer of the deep neural network can be set to 8, and the dimension of the corresponding input layer of the deep neural network is 8×2=18, where 2 is the initial state of the small convolutional code. Since the structure of the (2, 1, 2) convolutional code is relatively simple, one hidden layer is sufficient. Here, the size of the hidden layer is set to 512, and the activation function of the hidden layer is set to f(x)=relu( x). A deep neural network model is built according to the dimension of the output layer, the dimension of the input layer and the activation function of the hidden layer.

(2) Construct samples for the information code fields of noisy convolutional codes, and generate data sample sets conforming to the deep neural network model in batches.

Randomly generate an information sequence of length L, after encoding by (2, 1, 2) convolutional code, add noise with Gaussian white noise ranging from 1db to 7db to obtain a noisy convolutional code information of length 2×L sequence. Take L=1000, for the convenience of decoding, the last 7 bits can be set as zero information bits, so that the decoding process after the decoding of the eighth last bit ends, when constructing the sample, the noisy convolutional code information sequence according to the first The input dimension of the deep neural network model is segmented to form a noisy convolutional code information code field corresponding to the size of the deep neural network model. Take the information bit sliding window size as 8, and add Gaussian noise after encoding to obtain 16-bit codeword samples. The status bit of the first training sample starts from 00, followed by a 2×8 information sequence which is the information code field of the noisy convolutional code, which is used as the first training sample.

When taking the second training sample, the window slides one bit backward in the sequence direction in the information code field of the noisy convolutional code, and takes the second 0 of the status bit of the previous code field and the first bit of the code field as The status bit, plus the codeword bits added to the sliding window as input, the label is the onehot form of the code field before encoding, forming the second training sample, and so on, all the information sequences of length L are converted into corresponding The training sample set of the neural network with the input layer size of 16 and the input layer size of 8, the input contains the state bits of the code field and the code field, and the output is the onehot after the code field is decoded.

(3) After obtaining the data sample set, the hidden layer uses f(x)=relu(x) as the activation function, and uses the softmax classification method and the batch gradient descent method to train the deep neural network model. During the training of the deep neural network model, the optimal weights are obtained by two processes of feedforward calculation and backward propagation, which make the model have the ability to classify. The above steps are a complete training process. After training, the error will continue to decrease, that is to say, the deep neural network gradually learns to decode the noisy convolutional code information sequence. The training can be stopped by performing multiple trainings until the accuracy and error of the deep neural network stabilize. Here, the training times are selected as 2000 times.

(4) Input the convolutional code to be decoded into the trained deep neural network model for decoding, and complete the decoding of the convolutional code: randomly generate an information sequence for the trained deep neural network model, and go through (2 , 1, 2) After the convolutional code coding and the Gaussian white noise ranging from 1db to 7db are added to the noise, the obtained soft information with noise is input into the neural network according to the size of the input layer of the deep neural network. The initial state bit is 00. Taking the first information bit of the convolutional code as an example, the first training sample is decoded to obtain a decoding result with a length of 8. The bit is the first decoding result of the first information bit of the convolutional code. Taking the second training sample is equivalent to sliding one bit after the noisy codeword to get an output, and the first information bit of the convolutional code corresponds to the decoding output of the penultimate position on the second codeword. In this example, the sliding window size is 8. When the number of sliding times reaches 8, the output layer size of the deep neural network, 8 weak classifiers including this information bit are obtained. These 8 weak classifiers classify the information bits in different dimensions and have 8 decoding outputs, and use the integrated method to vote on the information bits to generate a strong classifier: that is, for the 8 weak classifiers 8 times of decoding outputs are counted, and a large number of decoding output results are obtained by voting as the decoding results of the information bits, so as to obtain better decoding performance for the information bits. The whole process is shown in Figure 3. The information bits of the same color are in the same position on the convolutional code, and each column is the decoding output of one sample, so it corresponds to the integration of 8 decoding results. Repeat this step for subsequent codewords, thus completing all decoding of the convolutional code.

Embodiment 2

The convolutional code decoding method based on the deep learning and ensemble method in this embodiment is as follows: the weak classifier and the number of weak classifiers are set, the weak classifier uses a perceptron to decode the convolutional code, and the deep neural network is used to decode the convolutional code. The depth of the network is set; finally, an ensemble method is used to vote the decoding result of the weak classifier to obtain the decoding output; the deep neural network is a fully connected neural network, a convolutional neural network, GAN or LSTM.

The above-mentioned embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, The simplification should be equivalent replacement manners, which are all included in the protection scope of the present invention.

Claims (4)

1. A convolutional code decoding method based on deep learning and integration method is characterized in that: the method comprises the following steps:

firstly, setting the number of weak classifiers and weak classifiers; determining model parameters of a deep neural network in a weak classifier, and establishing a deep neural network model; the deep neural network is a fully-connected neural network, a convolutional neural network, a GAN or an LSTM;

secondly, establishing a data sample set of convolutional code decoding;

thirdly, training a deep neural network model by using the data sample set in the second step and adopting a softmax classification mode and a batch gradient descent method;

fourthly, inputting the convolutional code to be decoded into the deep neural network model obtained in the third step for decoding; in the decoding process, a plurality of weak classifiers are obtained to classify and decode information bits corresponding to convolutional code information code segments to be decoded in different dimensions, an integrated method is adopted to vote for the decoded output of the information bits to generate a strong classifier so as to obtain final decoding, and the convolutional code decoding is completed;

the establishing of the data sample set of the convolutional code decoding refers to:

first, an information sequence of length L is randomly generated and passed through (n)₀，k₀M) after encoding the convolutional code, the length n is obtained by Gaussian white noise plus noise₀×L/k₀The information sequence of the noisy convolutional code;

secondly, adding 00 as a state bit in front of the noisy convolutional code information sequence, and segmenting the noisy convolutional code information sequence according to the input dimension of the deep neural network model in the first step to form a noisy convolutional code information code field corresponding to the size of the deep neural network model; wherein, for any one (n)₀，k₀M) the start states of the convolutional codes are all 00;

finally, carrying out sample construction on the noisy convolutional code information code word segment, and generating a data sample set conforming to the deep neural network model in batches;

the step of carrying out sample construction on the noisy convolutional code information code segment and generating a data sample set conforming to the deep neural network model in batches is as follows:

(1) in the information code field of the noisy convolutional code, front k₀X m bits are the status bits of the original codeword, the last n₀×n/k₀The bits are information code fields of the noisy convolutional codes and are used as first training samples;

(2) setting the size of a sample acquisition information bit window to be N, when a second training sample is taken, sliding a sample window backwards by one bit along the sequence direction for a code field of noisy convolutional code information, taking a second 0 of a state bit of a previous code field and a first bit of the previous code field as state bits, and adding code bits added in the sample window after sliding to serve as a second training sample;

(3) and in the same way, generating a data sample set conforming to the deep neural network model in batch according to the information code field of the full-section noisy convolutional code and the corresponding information bits.

2. The convolutional code decoding method based on deep learning and integration method as claimed in claim 1, wherein: in the first placeIn one step, the determining the model parameters of the deep neural network and establishing the deep neural network model includes: for any one (n)₀，k₀M) convolutional code, setting the output layer dimension of the deep neural network model as n and the input layer dimension as n₀×n/k₀(ii) a Setting an activation function of the hidden layer as f (x) ═ relu (x); and establishing a deep neural network model according to the output layer dimension, the input layer dimension and the activation function of the hidden layer.

3. The convolutional code decoding method based on deep learning and integration method as claimed in claim 1, wherein: in the third step, the optimal weight is obtained by adopting a mode of updating the weight in two processes of feedforward calculation and backward propagation during the training of the deep neural network model, so that the model has classification capability.

4. The convolutional code decoding method based on deep learning and integration method as claimed in claim 2, wherein: in the fourth step, the convolutional code to be decoded is input into the deep neural network model obtained in the third step for decoding; in the decoding process, a plurality of weak classifiers are obtained to classify and decode information bits corresponding to convolutional code information code segments to be decoded in different dimensions, an integrated method is adopted to vote for the decoded output of the information bits to generate a strong classifier so as to obtain final decoding, and the completion of convolutional code decoding means that:

(1) coding and denoising a convolutional code to be decoded to obtain a noisy convolutional code information sequence, and finally adding a zero information bit to an information bit of the noisy convolutional code information sequence and inputting the information bit into a deep neural network model;

(2) setting the initial state bit as 00, and decoding the first information bit;

(3) updating the state of the information sequence of the noisy convolutional code, sliding one bit behind the information sequence of the noisy convolutional code, and obtaining once decoding output after sliding one bit each time; when the number of sliding times behind the information sequence of the noisy convolutional code reaches the dimension size n of an output layer of the deep neural network model, acquiring n weak classifiers comprising the information bits, classifying the information bits on different dimensions by the n weak classifiers and outputting n decoding times; voting the information bit by adopting an integrated method to generate a strong classifier: counting the n decoding outputs, voting to obtain a large number of decoding outputs as the decoding result of the information bit;

(4) and repeating the decoding step for the subsequent noisy convolutional code information sequence to finish the convolutional code decoding.

Images