CN108880568A - A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks - Google Patents
A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks Download PDFInfo
- Publication number
- CN108880568A CN108880568A CN201810737117.6A CN201810737117A CN108880568A CN 108880568 A CN108880568 A CN 108880568A CN 201810737117 A CN201810737117 A CN 201810737117A CN 108880568 A CN108880568 A CN 108880568A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- decoding
- optimization method
- parameter optimization
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M13/00—Coding, decoding or code conversion, for error detection or error correction; Coding theory basic assumptions; Coding bounds; Error probability evaluation methods; Channel models; Simulation or testing of codes
- H03M13/03—Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words
- H03M13/23—Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words using convolutional codes, e.g. unit memory codes
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L1/00—Arrangements for detecting or preventing errors in the information received
- H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
- H04L1/0056—Systems characterized by the type of code used
- H04L1/0057—Block codes
Abstract
The present invention provides a kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks, since the method first collecting, arranging sample data;Then using sample data the characteristics of and size train network using back-propagation algorithm as according to being modeled;Then the likelihood ratio being calculated by reception signal is input to again and is completed in trained convolutional neural networks, export Q;L is finally initialized as Q, executes serial counteracting list decoding.This method avoids unnecessary calculating operation, to greatly reduce the decoding complexity of polarization code by the way that convolutional neural networks technology and polarization code decoding technique to be combined.
Description
Technical field
The invention belongs to field of communication technology, in particular to a kind of optimized with convolutional neural networks serial offset list and translates
The method of the polarization code decoding parameter of code algorithm.
Background technique
Polarization code is a kind of novel channel coding proposed by E.Arikan in 2008.Polarization code is that the first can pass through
Stringent mathematical method proves the constructivity encoding scheme for reaching channel capacity.It is serial to offset at the beginning of polarization code is suggested
(SC) decoding is also suggested therewith.SC decodes the path search process that can be considered to be on binary tree.SC decoding algorithm from
Code tree root node starts, and successively successively scans for leaf node layer, preferable from two subsequent middle selections after each layer of extension
One be extended.The characteristics of SC is decoded mainly has two aspects, is on the one hand that its complexity is low, decoding architecture is simple;It is another
Aspect is that it is theoretically proved to that shannon limit can be reached when code length is sufficiently large.But SC decoding algorithm is limited in code length
Under long configuration, error-correcting performance is undesirable.In order to improve performance, propose that serial offset list (SCL) is decoded.SCL decoding is SC
A kind of modified version of decoding.Unlike SC, SCL decoding algorithm be no longer from two it is subsequent it is middle selection it is preferable one into
Row extension, but retain the successor path for being not more than L item as much as possible, in next layer of extension, all this is not more than the time of L item
Routing diameter can all be extended respectively.When terminating the extension of leaf node layer, there is at most L path candidate to be retained in list
In.Since SCL decoding is only under higher signal-to-noise ratio, maximum-likelihood decoding performance can be realized, therefore cyclic redundancy check
(CRC) it is introduced for promoting the decoding performance of polarization code, this L path candidate is verified by using CRC, it is final defeated
Energy is by CRC's and most possible path candidate out.The SCL of CRC auxiliary has more better with LDPC code than Turbo code
Decoding performance, but as L increases, decoding complexity also be will increase.In order to solve this problem, adaptive serial counteracting list
(AD-SCL) decoding algorithm avoids calculating unnecessary path, greatly reduced decoding by adaptively controlling L
Complexity.
But under lower signal-to-noise ratio, AD-SCL can frequently occur high decoding complexity situation.AD-SCL algorithm is always
The initial value of L is configured to 1.If the AD-SCL decoding failure based on L=1, which can be updated to 2L by L and continue to translate
Code, until L=Lmax, LmaxIt is the maximum list size being set according to actual conditions.Under the configuration of low signal-to-noise ratio and L=1,
The probability of AD-SCL algorithm failure is high, it is therefore desirable to which frequent updating L value increases complexity.If when decoding beginning, by L
As soon as being initialized as a suitable value, executing primary decoding as much as possible can be successful, then, it is multiple that decoding will be significantly decreased
Miscellaneous degree.
In order to reduce decoding complexity, by finding a suitable L value under the premise of guaranteeing polarization code decoding performance
Calculation amount is reduced, the invention patent proposes a kind of serial counteracting list decoding parameter optimization side based on convolutional neural networks
Method optimizes L value by building and training convolutional neural networks, and to realize the target for reducing operand, it is multiple finally to reduce decoding
Miscellaneous degree.
Summary of the invention
The parameter optimization method for the convolutional neural networks auxiliary based on SCL decoding algorithm that the invention proposes a kind of, is being protected
In the case that card decoding performance is constant, decoding complexity is reduced by optimization L value, and this L value optimized is denoted as Q.
In the sample data preparation stage, different signal-to-noise ratio are based on, execute 100000 adaptive serial counteracting list decodings
Algorithm, the likelihood ratio being calculated by reception signal when will be successfully decoded each time and it is successfully decoded when corresponding L record,
The likelihood ratio that decoding success is recorded constitutes one group of sample data with corresponding L.60000 groups of sample datas are randomly selected,
75% group of data is randomly selected from this 60000 groups of data as training sample, and using remaining 25% group of data as test
Sample.
Convolutional neural networks (CNN) are a kind of feedforward neural networks and can be used as classifier, it is characterized in that hierarchical structure and
Training rules can be set according to the actual situation.When building CNN, hierarchical structure includes input, 2 convolutional layers, 2 pond layers
And 2 full articulamentums, network inputs are a likelihood ratio, regard likelihood ratio as a size be the matrix of 32*32;First
Convolutional layer has 6 filters, and the size of each filter is 5*5, and first convolutional layer maps out 6 characteristic patterns, each
The size of characteristic pattern is 28*28;First pond layer by obtaining 6 characteristic patterns after down-sampling, each characteristic pattern it is big
Small is 14*14;Second convolutional layer has 16 filters, and the size of each filter is 5*5, and second convolutional layer maps out
16 characteristic patterns, the size of each characteristic pattern are 10*10;Second pond layer is by obtaining 16 features after down-sampling
Figure, the size of each characteristic pattern are 5*5;The interstitial content of first full articulamentum is 50, the node of second full articulamentum
Number is 6;It is activation primitive by Relu function sets, label L includes 6 kinds of different sizes, i.e. L ∈ { 1,2,4,8,16,32 },
It is divided into 6 classes.Training CNN when, be based on supervised learning, train network using error backpropagation algorithm, by find out convolutional layer,
The error term of pond layer and full articulamentum biases to adjust weight and neuron, until completing the training of CNN.
End is decoded in polarization code, the likelihood ratio being calculated by reception signal is input in CNN model, obtains Q, and will
L is initialized as Q, executes serial counteracting list decoding.
In parameter optimisation procedure, it is applicable in following steps:
Step 1, prepare sample data, and sample data is pre-processed;
Step 2, convolutional neural networks, and training convolutional neural networks are built;
Step 3, the stage is decoded in polarization code, likelihood ratio is input in convolutional neural networks model, a value Q is obtained,
And L is initialized as Q, execute serial counteracting list decoding;
Wherein, prepare sample data in step 1 and refer to that execution 100000 times adaptive serial list decodings of offsetting are calculated
Method, the likelihood ratio being calculated by reception signal when will be successfully decoded each time and it is successfully decoded when corresponding L record, one
The likelihood ratio that secondary decoding success is recorded constitutes one group of sample data with corresponding L, randomly selects 60000 groups of sample datas, from
75% group of data is randomly selected in this 60000 groups of data as training sample, and using remaining 25% group of data as test specimens
This;Convolutional neural networks are built in step 2 refers to that the number of plies by convolutional layer, pond layer and full articulamentum is disposed as 2.
Beneficial effect
The present invention, which compares prior art, has following innovative point:
With convolutional neural networks come Optimal Parameters L.End is decoded in polarization code, likelihood ratio is input to, trained volume is completed
In product neural network, then exportable Q, realizes the Fast Classification of likelihood ratio.On this basis, in the premise for guaranteeing polarization code performance
Under, decoding algorithm can reduce decoding complexity with suitable L value.
Convolutional neural networks technology and polarization code decoding technique are combined.In the sample data preparation stage, sample number
Adaptive serial counteracting list decoding is performed a plurality of times according to deriving from;It is building and training convolutional neural networks stage, network
Input be likelihood ratio, the output of network is L;At decoding end, it is based on SCL decoding algorithm, L is initialized as Q.At this point, SCL is translated
Successfully probability is higher for code, does not need frequent updating L value, avoids unnecessary arithmetic operation, reduces the operand of decoder, from
And computation complexity is greatly reduced.
Detailed description of the invention
The convolutional neural networks schematic diagram that Fig. 1 is built when being determining convolutional neural networks structure;
Fig. 2 is the method flow diagram of Optimal Parameters L.
Specific embodiment
Below in conjunction with drawings and examples, the present invention will be further described.
The present invention provides a kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks, mainly includes
Prepare sample data, build simultaneously training convolutional neural networks and decoding three parts.In the sample data preparation stage, first in difference
Executed under signal-to-noise ratio 100000 times it is adaptive serial offset list decodings, and will be successfully decoded each time when by reception signal
The likelihood ratio that is calculated and it is successfully decoded when corresponding L record, then randomly select 60000 groups of sample datas, finally
75% group of data is randomly selected from this 60000 groups of data as training sample, and using remaining 25% group of data as test
Sample;Build with the training convolutional neural networks stage, first according to the characteristics of training sample and size determines the layer of network
Level structure and parameter, build convolutional neural networks, find out the error term of convolutional layer, pond layer and full articulamentum then to adjust
Weight and neuron biasing;End is decoded in polarization code, the likelihood ratio being calculated by reception signal is input to CNN model first
In, Q is then obtained, L is finally initialized as Q, executes serial counteracting list decoding.
It executes 100000 adaptive serial list decodings of offsetting in different signal-to-noise ratio in the sample data preparation stage and calculates
Method, and will be successfully decoded each time when the likelihood ratio that is calculated by reception signal and it is successfully decoded when corresponding L record,
The likelihood ratio that decoding success is recorded constitutes one group of sample data with corresponding L;Then 60000 groups of sample numbers are randomly selected
According to;75% group of data is finally randomly selected from this 60000 groups of data as training sample, and remaining 25% group of data are made
For test sample, sample data is converted to by numeric data using normalized.The present embodiment adaptively will serially offset column
The code length of table decoding algorithm is set as 1024, and code rate is set as 0.5, Lmax=32, CRC length are 16, by number of training and survey
Examination sample number is respectively set to 45000 and 15000.
It is building with the training convolutional neural networks stage, the convolutional neural networks built are as shown in Figure 1, include input, 2
Convolutional layer, 2 pond layers and 2 full articulamentums, network inputs are a likelihood ratio, regard likelihood ratio as a size be
The matrix of 32*32;First convolutional layer has 6 filters, and the size of each filter is 5*5, first convolutional layer mapping
6 characteristic patterns out, the size of each characteristic pattern are 28*28;First pond layer is by obtaining 6 features after down-sampling
Figure, the size of each characteristic pattern are 14*14;Second convolutional layer has 16 filters, and the size of each filter is 5*
5, second convolutional layer maps out 16 characteristic patterns, and the size of each characteristic pattern is 10*10;Second pond layer is adopted under passing through
16 characteristic patterns are obtained after sample, the size of each characteristic pattern is 5*5;The interstitial content of first full articulamentum is 50;L is
Label, each sample correspond to a kind of label, and the label of the present embodiment has 6 kinds of different sizes, i.e. L=1, L=2, L=4,
L=8, L=16, L=32 can be classified as 6 classes, therefore set 6 for the interstitial content of second full articulamentum.The present embodiment
Relu function is set by activation primitive, using the pond Max Pooling method, calculates convolutional layer, pond layer and full articulamentum
Output.Based on supervised learning, network is trained using back-propagation algorithm, the present embodiment connects all second, network entirely
The error sum of squares of node layer is as objective function;On this basis, with stochastic gradient descent optimization method to objective function
It optimizes;By optimization after, respectively obtain the error term of convolutional layer, the error term of pond layer, the error term of full articulamentum,
The update method of weight and the update mode of bias term.According to obtained training rules, the weight of continuous corrective networks and partially
Item is set, until completing all sample trainings.
End is decoded in polarization code, comes Optimal Parameters L, optimization method flow chart such as Fig. 2 of parameter L by convolutional neural networks
It is shown.The likelihood ratio being calculated by reception signal is input to first and is completed in trained convolutional neural networks, the likelihood
The representation of ratio is the matrix of a 32*32, exports Q;L is initialized as Q again, executes serial counteracting list decoding;
Then CRC check is carried out to L item decoding path candidate, if there is one or the path candidate of more than one is by CRC check, then
Decoding is simultaneously exited in one most possible path of output, otherwise, L is updated to 2L;It completes after updating, judges whether L is greater than
LmaxIf being not more than Lmax, then continue serially to offset list decoding, otherwise, exit decoding.The present embodiment will serially offset column
The code length of table decoding is set as 1024, and code rate is set as 0.5, Lmax=32, CRC length are 16.
The above description is merely a specific embodiment, but the scope of protection of the present invention is not limited thereto, any ripe
Those skilled in the art are known in technical scope proposed by the present invention, the variation or replacement that can be readily occurred in all are answered
This is included within the scope of the present invention.
Claims (3)
1. a kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks, which is characterized in that the method
Optimize the serial list size L for offsetting list decoding with convolutional neural networks, the parameter optimization method includes the following steps:
Step 1, prepare sample data, and sample data is pre-processed;
Step 2, convolutional neural networks, and training convolutional neural networks are built;
Step 3, the stage is decoded in polarization code, likelihood ratio is input in convolutional neural networks model, obtains a value Q, and by L
It is initialized as Q, executes serial counteracting list decoding.
2. a kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks according to claim 1,
It is characterized in that, preparing sample data in step 1 refers to execution 100000 times adaptive serial counteracting list decodings,
The likelihood ratio that is calculated when will be successfully decoded each time by reception signal and it is successfully decoded when corresponding L record, once
The likelihood ratio that decoding success is recorded constitutes one group of sample data with corresponding L, 60000 groups of sample datas is randomly selected, from this
75% group of data is randomly selected in 60000 groups of data as training sample, and using remaining 25% group of data as test sample.
3. a kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks according to claim 1,
It is characterized in that, building convolutional neural networks in step 2 refers to that the number of plies by convolutional layer, pond layer and full articulamentum is all provided with
It is set to 2.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810737117.6A CN108880568A (en) | 2018-07-06 | 2018-07-06 | A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810737117.6A CN108880568A (en) | 2018-07-06 | 2018-07-06 | A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108880568A true CN108880568A (en) | 2018-11-23 |
Family
ID=64299538
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810737117.6A Pending CN108880568A (en) | 2018-07-06 | 2018-07-06 | A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108880568A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109450459A (en) * | 2019-01-16 | 2019-03-08 | 中国计量大学 | A kind of polarization code FNSC decoder based on deep learning |
CN111313914A (en) * | 2019-11-05 | 2020-06-19 | 北京航空航天大学 | SCL simplified decoding method based on neural network classifier |
CN114073024A (en) * | 2019-06-03 | 2022-02-18 | 加利福尼亚大学董事会 | Convolutional precoding and decoding of polar codes |
WO2022083400A1 (en) * | 2020-10-20 | 2022-04-28 | 华为技术有限公司 | Decoding method and apparatus |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106295682A (en) * | 2016-08-02 | 2017-01-04 | 厦门美图之家科技有限公司 | A kind of judge the method for the picture quality factor, device and calculating equipment |
CN106506009A (en) * | 2016-10-31 | 2017-03-15 | 中国石油大学(华东) | A kind of interpretation method of polarization code |
CN108140131A (en) * | 2015-10-04 | 2018-06-08 | 艾腾怀斯股份有限公司 | For convolutional network to be applied to the system and method for spatial data |
US20190340497A1 (en) * | 2016-12-09 | 2019-11-07 | William Marsh Rice University | Signal Recovery Via Deep Convolutional Networks |
-
2018
- 2018-07-06 CN CN201810737117.6A patent/CN108880568A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108140131A (en) * | 2015-10-04 | 2018-06-08 | 艾腾怀斯股份有限公司 | For convolutional network to be applied to the system and method for spatial data |
CN106295682A (en) * | 2016-08-02 | 2017-01-04 | 厦门美图之家科技有限公司 | A kind of judge the method for the picture quality factor, device and calculating equipment |
CN106506009A (en) * | 2016-10-31 | 2017-03-15 | 中国石油大学(华东) | A kind of interpretation method of polarization code |
US20190340497A1 (en) * | 2016-12-09 | 2019-11-07 | William Marsh Rice University | Signal Recovery Via Deep Convolutional Networks |
Non-Patent Citations (1)
Title |
---|
王玉环等: "基于神经网络的信道译码算法研究综述", 《中国传媒大学学报》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109450459A (en) * | 2019-01-16 | 2019-03-08 | 中国计量大学 | A kind of polarization code FNSC decoder based on deep learning |
CN114073024A (en) * | 2019-06-03 | 2022-02-18 | 加利福尼亚大学董事会 | Convolutional precoding and decoding of polar codes |
CN114073024B (en) * | 2019-06-03 | 2024-04-26 | 加利福尼亚大学董事会 | Convolutional precoding and decoding of polar codes |
CN111313914A (en) * | 2019-11-05 | 2020-06-19 | 北京航空航天大学 | SCL simplified decoding method based on neural network classifier |
CN111313914B (en) * | 2019-11-05 | 2021-09-28 | 北京航空航天大学 | SCL simplified decoding method based on neural network classifier |
WO2022083400A1 (en) * | 2020-10-20 | 2022-04-28 | 华为技术有限公司 | Decoding method and apparatus |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108880568A (en) | A kind of serial counteracting list decoding parameter optimization method based on convolutional neural networks | |
CN108777584A (en) | A kind of fast Optimization of polarization code decoding parameter | |
Ozturk et al. | Hybrid artificial bee colony algorithm for neural network training | |
Zhou et al. | M2kd: Multi-model and multi-level knowledge distillation for incremental learning | |
Pujol et al. | Evolving the topology and the weights of neural networks using a dual representation | |
CN103929210B (en) | Hard decision decoding method based on genetic algorithm and neural network | |
CN111078911B (en) | Unsupervised hashing method based on self-encoder | |
CN107517095A (en) | A kind of polarization code coding/decoding method of unequal piece-wise verification | |
CN109245776A (en) | A kind of polarization code decoding parameter determination method based on deep neural network | |
CN109450459B (en) | Polarization code FNSC decoder based on deep learning | |
CN107067077A (en) | A kind of weighting algorithm of convolutional neural networks | |
CN109740695A (en) | Image-recognizing method based on adaptive full convolution attention network | |
CN110138390A (en) | A kind of polarization code SSCL algorithm decoder based on deep learning | |
CN113850385A (en) | Coarse and fine granularity combined neural network pruning method | |
Aly et al. | Derivative-free optimization of neural networks using local search | |
CN108964672A (en) | A kind of polarization code coding method based on deep neural network | |
CN113158580A (en) | One-dimensional stock layout method for solid wood board | |
CN106452675B (en) | A kind of spherical decoding method of polarization code | |
CN110830167A (en) | Multi-bit flipping-based polar code Fast-SSC decoding algorithm | |
Xia et al. | Improving the performance of stock trend prediction by applying GA to feature selection | |
CN108733806A (en) | A kind of data processing method of multi-level convolutional neural networks | |
CN105515588B (en) | A kind of LDPC CC high speed decoders | |
da Silva et al. | A multi-objective grammatical evolution framework to generate convolutional neural network architectures | |
CN116318551A (en) | Intermediate channel selection and decoding method of LDPC-Polar cascade system | |
CN112577744B (en) | Rolling bearing fault identification method based on combination of SPA-map and ResNet |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181123 |