WO2020125421A1 - Scma系统的dnn解码方法及解码通信设备 - Google Patents

Scma系统的dnn解码方法及解码通信设备 Download PDF

Info

Publication number
WO2020125421A1
WO2020125421A1 PCT/CN2019/123144 CN2019123144W WO2020125421A1 WO 2020125421 A1 WO2020125421 A1 WO 2020125421A1 CN 2019123144 W CN2019123144 W CN 2019123144W WO 2020125421 A1 WO2020125421 A1 WO 2020125421A1
Authority
WO
WIPO (PCT)
Prior art keywords
scma
output
layer
signal
decoding
Prior art date
Application number
PCT/CN2019/123144
Other languages
English (en)
French (fr)
Inventor
林进挚
赵希敏
胡金星
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2020125421A1 publication Critical patent/WO2020125421A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00Arrangements for detecting or preventing errors in the information received

Definitions

  • the invention relates to the field of wireless communication, in particular to a DNN decoding method of SCMA system and decoding communication equipment.
  • SCMA Separatse Code Multiple Access
  • SCMA sparse code multiple access
  • MPA message passing Algorithm
  • ML Maximum Likelihood
  • the present invention provides a deep neural network (DNN)-based SCMA signal decoding method and decoding communication device, which can greatly reduce the complexity of the decoder, and at the same time it is easy to operate, by This improves decoding efficiency and performance.
  • DNN deep neural network
  • the present invention provides the following technical solutions:
  • the step S1 includes:
  • the SCMA receiver receives the SCMA signal in the physical environment, and records and stores the SCMA signal data
  • the step S2 includes:
  • y l-1 is the data output of the previous layer
  • b l are the weight and offset of this layer respectively;
  • W d , b d respectively represent the weight and offset set of the entire decoder;
  • d( ⁇ ; W d , b d ) represents the final output value of the decoder, and its value is the decoded bit vector
  • the output probability ⁇ (x) of the output node calculated by formula (2), [0,1] is the value range, and
  • ⁇ i [ ⁇ ] represents the value of the i-th element of the vector.
  • each resource block is provided with two input nodes, respectively corresponding to the real part and the imaginary part of the complex signal.
  • a random gradient descent method is used to train the above SCMA decoder model in step S3, so that the cross-entropy loss function satisfies the condition described in the following formula (4);
  • step S3 further includes:
  • Var[ ⁇ ] means taking the variance
  • y l-1 is the initial state of the output of layer l-1
  • W l is the weight of layer l
  • N l,i is the total number of input nodes of layer l;
  • the step S3.2 includes:
  • Z l, k (i) represents The k-th element in the vector
  • Nb indicates that the complete training set is divided into Nb batches
  • i indicates the i-th batch
  • the step S4 includes:
  • S4.2 Receive the signal sent by the SCMA transmitter, and transfer the received SCMA signal to the loaded SCMA decoder model for decoding processing, and real-time online decoding output after compilation and operation.
  • a decoding communication device for implementing the above decoding method is also provided, which includes:
  • Training data generation module which is used to associate the source code word with the SCMA signal data, and obtain the signal data set under different signal-to-noise ratio conditions according to the correlation result data, thereby obtaining the training data set;
  • Model generation module which is used to build the SCMA decoder model based on deep neural network
  • Model training into a module which is used to train the above SCMA decoder model, and save the SCMA decoder model after the training is completed;
  • the development board platform is used to load the optimized decoder model file, and load the received SCMA signal into the decoder model file for real-time online decoding processing, and output the processing result.
  • the present invention adopts a software radio platform to build an SCMA system, and at the same time builds an SCMA decoder model, and deploys the SCMA decoder model to development boards such as AIR-T.
  • the operation process is simple and fast, especially It is suitable for overloaded non-orthogonal multiple access wireless communication system, which can improve the accuracy of SCMA decoding without increasing the complexity of SCMA decoder.
  • the present invention The DMA-based SCMA decoder has improved performance in terms of computational complexity and decoding error rate.
  • FIG. 2 is a schematic structural diagram of a decoding device in Embodiment 2;
  • FIG. 3 is a schematic structural diagram of a model generation module in Embodiment 2;
  • FIG. 4 is a schematic structural diagram of a model training module in Embodiment 2;
  • FIG. 5 is a schematic structural diagram of a development board platform in Embodiment 2.
  • This embodiment provides a DNN decoding method of an SCMA system, which includes the following steps:
  • step S1 includes the following steps:
  • the setting related parameters include: setting the UHD sink module parameters of USRP B210, such as setting the center frequency to 500 MHz, Gain Value to 50, Gain Type to absolute, Antenna to TX/RX, sampling rate to 100K, and others
  • the parameters are all default parameters, etc.; set the USRP source module parameters (see USRP B210 UHD sink module parameter setting method); and set some other module parameters, including the transmitted codeword, packet header format, packet header check, and carrier allocation parameters (including FFT length, carrier, preamble carrier, preamble code word, synchronization symbol, Cyclic Prefixer length, one or several items in the compilation operation; the specific setting method of the above parameters depends on the actual design needs, and is not specifically limited here ;
  • the SCMA receiver receives the SCMA signal in the physical environment, and records and stores the SCMA signal data
  • the SCMA decoder of the present invention can be adapted to signals with different signal-to-noise ratios, and has a more comprehensive generalization capability, while avoiding the decoder from appearing excessively. Fitting and underfitting;
  • each resource block has two input nodes, corresponding to the real and imaginary parts of the complex signal, as shown in Figure 1, where the input nodes Ri and Ii correspond to the real and imaginary parts of the SCMA signal, respectively.
  • the total number of nodes in the input layer is 2*K (where K is the number of resource blocks);
  • iNj represents the j-th node of the i-th hidden layer
  • y l-1 is the data output of the previous layer
  • b l are the weight and offset of this layer respectively; it should be noted that there can be several hidden layers. If “this layer” is not the first hidden layer, then “upper layer” is the hidden layer's The last hidden layer, as shown in Figure 1, if “this layer” is a hidden layer of 3N1, then “upper layer” is a hidden layer of 2N1, and so on; if “this layer” is the first hidden layer Layer, the “upper layer” is the input layer, as shown in Figure 1, if the "this layer” is a hidden layer of 1N1, the “upper layer” is the input layer, and so on;
  • W d , b d respectively represent the weight and offset set of the entire decoder;
  • d( ⁇ ; W d , b d ) represents the final output value of the decoder, and its value is the decoded bit vector
  • the output probability ⁇ (x) of the output node calculated by formula (2), [0,1] is the value range, and ⁇ i [ ⁇ ] represents the value of the i-th element of the vector;
  • the SCMA decoder model in this embodiment is a DNN-based model, including an input layer, 6 hidden layers, and an output layer, among which the input layer, hidden layer, and Refer to steps S2.1-2.4 for the establishment method of the output layer, which will not be repeated here;
  • the commonly used stochastic gradient descent method is used to train the above SCMA decoder model to search for the optimal weights and offsets that approximate each hidden layer, so that the cross-entropy loss function satisfies the conditions described in the following formula (4);
  • the ADAM algorithm is used for SCMA decoder model training in this step, thereby facilitating direct use in artificial intelligence learning systems such as Tensorflow;
  • the training data set is divided into several batches of data sets and sent to the above model for training, and the linear output of each layer is recorded for then Is the linear output set of the complete training data set at this layer, where N b is the size of the batch data set;
  • variable Xavier initialization method and variable batch normalization method are also introduced in this step; specifically, it includes the following steps:
  • the S3.1 variable Xavier is initialized so that the initial state y l output by each layer (which can be any layer of the input layer, hidden layer, and output layer) needs to satisfy the conditions described in the following formula (5):
  • Var[ ⁇ ] means taking the variance
  • y l-1 is the initial state of the output of layer l-1
  • W l is the weight of layer l
  • N l,i is the total number of input nodes of layer l, and for the forward direction
  • N l,i is the total number of input nodes of layer l
  • N l,o is the output of layer l Total number of nodes; preferably, in order to avoid the above two conditions from conflicting with each other, this embodiment selects Randomly initialize the weights;
  • Z l, k (i) represents The k-th element in the vector
  • Nb indicates that the complete training set is divided into Nb batches
  • i indicates the i-th batch
  • is a constant (a trainable constant whose value is determined by the training result);
  • This embodiment also provides an SCMA decoding device for implementing the decoding method in Embodiment 1, as shown in FIG. 2, which includes:
  • the signal transmitter 1 is used to send the SCMA signal generated in the SCMA system to the physical environment;
  • the signal receiver 2 is used to receive the SCMA signal in the physical environment, and record and store the SCMA signal data;
  • the training data generation module 3 is used to associate the source code word of the signal transmitter 1 with the SCMA signal data, record and store the correlation result data, and obtain the signal data set under different signal-to-noise ratio conditions according to the correlation result data, thereby Obtain a training data set, and load and store the training data set into a Numpy array;
  • Model generation module 4 which is used to build a SCMA decoder model based on a deep neural network; specifically, as shown in FIG. 3, it includes: an input layer building module 41, which is used to receive SCMA signal data and has at least one resource Block; hidden layer establishment module 42, which is used for SCMA signal feature learning and extraction, and the data output of the hidden layer is completed according to formula (1); output layer establishment module 43, which is used to calculate at least according to formula (2) The output probability of an output layer node, and the degree of inconsistency between the predicted value and the true value is calculated according to the formula (3), and the final output includes the decoding result of the codeword of at least one user;
  • the model is trained into module 5, which is used to train the above SCMA decoder model using the commonly used stochastic gradient descent method to search for the optimal weights and offsets that approximate each hidden layer, so that the cross-entropy loss function satisfies formula (4)
  • the model training module 5 further includes: a variable initialization module 51, which is used to make the initial state y output by each layer l Both need to satisfy the conditions described in formula (5); and the variable normalization module 52, which is used to batch normalize the data of each layer before nonlinear activation;
  • Model file optimization module 6 which is used to export the SCMA decoder model to the UFF file format, and then use the AI inference platform such as TensorRT to optimize the SCMA decoder model in the UFF file format into a .plan file;
  • Development board platform (such as AIR-T, etc.) 7, as shown in Figure 5, which includes an import module 71 (such as TensorRT GRC module, etc.) and a radio receiving module 72, used to load the optimized decoder model file, and will receive The received SCMA signal is loaded into the decoder model file for real-time online decoding processing, and the processing result is output.
  • an import module 71 such as TensorRT GRC module, etc.
  • radio receiving module 72 used to load the optimized decoder model file, and will receive The received SCMA signal is loaded into the decoder model file for real-time online decoding processing, and the processing result is output.
  • This embodiment provides a wireless communication device, which includes the decoding device described in Embodiment 3.
  • the traditional SCMA decoding method mostly uses the MPA (Message Pass Algorithm) algorithm, and its complexity is O(X df ), where d f is the overload degree of the user, and the decoder in the present invention is a DNN-based decoder, Its complexity is (See complexity comparison in Table 1), where N L is the number of hidden layers, and N HN is the number of hidden layer nodes. It can be seen from the calculation process that the complexity of the SCMA decoder based on the traditional MPA algorithm will increase exponentially with the increase of the number of users, while the complexity of the SCMA decoder based on DNN increases slowly.
  • MPA Message Pass Algorithm
  • Table 2 shows the comparison of the complexity of the SCMA decoder based on the MPA algorithm and the SCMA decoder based on the DNN of the present invention when the MPA algorithm is iterated 5 times and the signal-to-noise ratio and bit error rate are the same. It can be seen from this that in this particular example, the complexity of the SCMA decoder based on the DNN of the present invention is improved by nearly 50% compared to the MPA algorithm.
  • the present invention provides a SCMA decoder based on a deep neural network, which uses a software radio platform to build an SCMA system, at the same time build an SCMA decoder model, and deploy the SCMA decoder model to AIR-T, etc.
  • its operation process is simple and fast, especially suitable for overloaded non-orthogonal multiple access wireless communication system, which can improve the accuracy of SCMA decoding without increasing the complexity of SCMA decoder, compared with the traditional
  • the SCMA decoder based on the DNN in the present invention has improved performance in terms of computational complexity and decoding error rate.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Error Detection And Correction (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本发明公开了一种SCMA系统的DNN解码方法及解码通信设备,其包括S1、搭建SCMA系统以及获得训练样本数据集;S2、建立基于深度神经网络的SCMA解码器模型;S3、训练上述SCMA解码器模型;以及S4、部署SCMA解码器模型,并通过所述SCMA解码器模型对SCMA信号进行解码。本发明可在不增加SCMA解码器复杂度的前提下提高SCMA解码准确率,相对于传统基于MPA算法的SCMA解码器而言,本发明中基于DNN的SCMA解码器在计算复杂度与解码误码率方面性能都有所提升。

Description

SCMA系统的DNN解码方法及解码通信设备 技术领域
本发明涉及无线通信领域,具体为一种SCMA系统的DNN解码方法及解码通信设备。
背景技术
传统SCMA(Sparse Code Multiple Access,SCMA,稀疏码多址接入)解码器采用消息传递算法(Message Passing Algorithm,MPA),结合先验概率,利用因子图在用户节点和资源节点之间迭代更新后验概率消息,以尽可能准确地解析原多用户发送的码字。相比于最大似然算法(Maximum Likelihood,ML)检测,MPA解码器的算法复杂度虽有所降低,但硬件实现依然较为困难,其复杂度随用户数量成指数级增长,从而导致解码效率低,不能满足未来5G系统的部署需求。
发明内容
解决的技术问题
针对现有技术的不足,本发明提供了一种基于深度神经网络(Deep Neural Network,DNN)的SCMA信号解码方法及解码通信设备,其可以大幅降低解码器的复杂度,同时其便于操作,由此提高解码效率和性能。
技术方案
为实现上述目的,本发明提供如下技术方案:
一方面,提出一种SCMA系统的DNN解码方法,其包括如下步骤:
S1、搭建用于产生SCMA信号的SCMA系统,以及在将SCMA信号 发送装置的源码字与SCMA信号接收装置接收到的SCMA信号数据进行关联后获得训练样本数据集;
S2、建立基于深度神经网络的SCMA解码器模型;
S3、根据所述训练样本数据集训练上述SCMA解码器模型;
以及S4、将训练后的SCMA解码器模型加载至解码平台,并通过所述SCMA解码器模型对SCMA信号进行解码。
优选的,所述步骤S1包括:
S1.1、搭建用于产生SCMA信号的SCMA系统,以及搭建SCMA发送机和SCMA接收机,记录并存储SCMA发送机的源码字,并通过SCMA发送机将所述SCMA系统产生的SCMA信号发送至物理环境中;
S1.2、SCMA接收机接收物理环境中的SCMA信号,且记录并存储SCMA信号数据;
S1.3、将SCMA发送机的源码字与SCMA接收机存储的SCMA信号数据进行关联,记录并存储关联结果数据;
S1.4、重复步骤S1.1-1.3,记录并存储每次重复得到的关联结果数据,以此得到不同信噪比条件下的信号数据集;
以及S1.5、整理不同信噪比条件下的信号数据集中的源码字以及与其关联的SCMA信号数据,由此得到所述训练数据集。
优选的,所述步骤S2包括:
S2.1、建立解码器的输入层,且所述输入层包括至少一个用于接收SCMA信号数据的资源块;
S2.2、建立解码器的隐藏层,并通过下述公式(1)完成所述隐藏层的数据输出,
Figure PCTCN2019123144-appb-000001
其中,y l-1是上一层的数据输出,
Figure PCTCN2019123144-appb-000002
和b l分别是本层的权重和偏置;
S2.3、建立解码器的输出层,且所述输出层用于输出包括至少一个用户的码字的解码结果;每个用户对应m=log 2(M)个输出层节点,分别对应解码出来的m个比特向量,所述输出层的节点总数是mJ,其中M是每个用户的码字个数,J是用户数;
S2.4、按照下述公式(2)计算至少一个输出层节点的输出概率σ(x):
σ(x)=(1+e -x) -1     (2);
以及按照下述公式(3)计算预测值与真实值之间的不一致程度:
Figure PCTCN2019123144-appb-000003
其中,y=(y 1,…,y 2K) T是上述训练数据集中从物理环境中收集的SCMA信号数据;b=(b 1,…,b mJ) T是与之关联的源码字;W d,b d分别代表整个解码器的权重和偏置集合;d(·;W d,b d)表示解码器的最后输出值,其值是解码出来的比特向量
Figure PCTCN2019123144-appb-000004
此处的
Figure PCTCN2019123144-appb-000005
Figure PCTCN2019123144-appb-000006
也即公式(2)计算出的输出节点的输出概率σ(x),[0,1]为取值范围,π i[·]表示取向量的第i个元素的值。
优选的,所述步骤S2.1中,每个资源块设有两个输入节点,分别对应复数信号的实部和虚部。
优选的,步骤S3中采用随机梯度下降法对上述SCMA解码器模 型进行训练,使得交叉熵损失函数满足下述公式(4)所述条件;
Figure PCTCN2019123144-appb-000007
优选的,步骤S3中还包括:
S3.1变量Xavier初始化,使得每一层输出的初始状态y l均需要满足以下公式(5)所述条件:
Var[y l]=N l,i·Var[W l]·Var[y l-1]       (5);
其中,Var[·]表示取方差;y l-1是第l-1层输出的初始状态;W l是第l层的权重;N l,i是第l层输入节点总数;
S3.2、将每一层的数据在完成线性输出前进行批量归一化;
S3.3按照步骤S3.1和3.2反复训练SCMA解码器模型,且在训练完成后保存SCMA解码器模型。
优选的,步骤S3.1中,
Figure PCTCN2019123144-appb-000008
其中,N l,o是第l层输出节点总数,且对于反向梯度传递而言,要求N l,o·Var[W l]=1。
优选的,所述步骤S3.2包括:
S3.2.1、对于每个输出节点k∈{1,…,N l,0},分别按照公式(6)、(7)计算均值μ β,k和方差σ β,k 2
Figure PCTCN2019123144-appb-000009
Figure PCTCN2019123144-appb-000010
其中,Z l,k (i)表示
Figure PCTCN2019123144-appb-000011
的向量中的第k个元素,Nb表示完整训练集被分成Nb个批次,i表示第i个批次;
S3.2.2、按照公式(8)进行标准化运算:
Figure PCTCN2019123144-appb-000012
其中,
Figure PCTCN2019123144-appb-000013
S3.2.3、按照公式(9)进行缩放与转换运算,以获得批量归一化结果a l,k i
Figure PCTCN2019123144-appb-000014
其中,
Figure PCTCN2019123144-appb-000015
以及β l,k i是在训练期间与原权重和偏置一起学习得到的缩放转换系数,最终将a l,k i代入公式(1)中进行非线性计算。
优选的,所述步骤S4包括:
S4.1、将经步骤S3保存的SCMA解码器模型导出并优化,再将其加载至开发板平台;
S4.2、接收SCMA发送机发送的信号,并将收到的SCMA信号转至已加载的SCMA解码器模型进行解码处理,编译运行后实时在线解码输出。
另一方面,还提供一种用于实现上述解码方法的解码通信设备,其包括:
训练数据生成模块,其用于将源码字与SCMA信号数据进行关联,且根据关联结果数据得到不同信噪比条件下的信号数据集,由此得到训练数据集;
模型生成模块,其用于建立基于深度神经网络的SCMA解码器模 型;
模型训练成模块,其用于对上述SCMA解码器模型进行训练,且在训练完成后保存SCMA解码器模型;
开发板平台,其用于加载优化后的解码器模型文件,且将接收到的SCMA信号加载至所述解码器模型文件中进行实时在线解码处理,并输出处理结果。
有益效果
与现有技术相比,本发明采用软件无线电平台搭建SCMA系统,同时构建SCMA解码器模型,并将所述SCMA解码器模型部署至AIR-T等开发板上,其操作过程简单、快捷,尤其适用于过载的非正交多址接入无线通信系统,由此可在不增加SCMA解码器复杂度的前提下提高SCMA解码准确率,相对于传统基于MPA算法的SCMA解码器而言,本发明中基于DNN的SCMA解码器在计算复杂度与解码误码率方面性能都有所提升。
附图说明
图1为实施例一中SCMA解码器模型的网络结构图;
图2为实施例二中解码设备的结构示意图;
图3为实施例二中模型生成模块的结构示意图;
图4为实施例二中模型训练成模块的结构示意图;
图5为实施例二中开发板平台的结构示意图。
具体实施方式
下面将结合本发明实施例中的附图,对本发明实施例中的技术 方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本发明保护的范围。
实施例一:
本实施例提供了一种SCMA系统的DNN解码方法,其包括如下步骤:
S1、搭建用于产生SCMA信号的SCMA系统,以及在将SCMA信号发送装置的源码字与SCMA信号接收装置接收到的SCMA信号数据进行关联后获得训练样本数据集;
具体的,所述步骤S1包括如下步骤:
S1.1、利用软件无线电结构平台(如GNU Radio等)搭建用于产生SCMA信号的SCMA系统,以及利用软件无线电设备(包括USRP B210)等搭建SCMA发送机和SCMA接收机,设置相关参数,记录并存储SCMA发送机的源码字,并通过SCMA发送机将将所述SCMA系统产生的SCMA信号发送至物理环境中;
所述设置相关参数包括:设置USRP B210的UHD sink模块参数,如将中心频率设为500MHZ、Gain Value设为50、Gain Type设为absolute、Antenna设置为TX/RX、采样率设为100K、其他参数均为默认参数等;设置USRP source模块参数(参见USRP B210的UHD sink模块参数设置方式);以及设置其他一些模块参数,包括发送的码字、包头格式、包头校验以及载波分配参数(包含FFT长度、载波、前导码载波、前导码字、同步符号、Cyclic Prefixer长度、编 译运行中的一项或几项);上述参数的具体设置方式根据实际设计需要而定,在此不做具体限定;
S1.2、SCMA接收机接收物理环境中的SCMA信号,且记录并存储SCMA信号数据;
S1.3、将SCMA发送机的源码字与SCMA接收机存储的SCMA信号数据进行关联,记录并存储关联结果数据;
S1.4、重复步骤S1.1-1.3,且每次重复时对应更改Gain Value的值,记录并存储每次重复得到的关联结果数据,以此得到不同信噪比条件下的信号数据集;由此,在收集数据的过程中,通过更改发送机和接收机的增益,可使得本发明的SCMA解码器适应不同信噪比的信号,具有更全的泛化能力,同时避免解码器出现过拟合和欠拟合;
S1.5、整理不同信噪比条件下的信号数据集中的源码字以及与其关联的SCMA信号数据,由此得到训练数据集,且将所述训练数据集加载并存储至Numpy数组中;
S2、建立基于深度神经网络(Deep Neural Network,DNN)的SCMA解码器模型(如图1所示);其具体包括如下步骤:
S2.1、建立解码器的输入层,且所述输入层包括至少一个用于接收SCMA信号数据的资源块;考虑到接收到的信号是幅度相位(IQ)组成的复数信号,因此在所述输入层中,每个资源块设有两个输入节点,分别对应复数信号的实部和虚部,如图1所示,其中输入节点Ri和Ii分别对应SCMA信号的实部和虚部,如此,所述输入层的节点总数是2*K个(其中K是资源块个数);
S2.2、建立解码器的隐藏层,用于进行SCMA信号特征学习与提取,图1中iNj代表第i隐藏层的第j个节点;
进一步的,通过下述公式(1)(也即Tanh函数)完成所述隐藏层的数据输出,
Figure PCTCN2019123144-appb-000016
其中,y l-1是上一层的数据输出,
Figure PCTCN2019123144-appb-000017
和b l分别是本层的权重和偏置;需要说明的是,所述隐藏层可以有若干个,若“本层”不是第一个隐藏层,则“上一层”为该隐藏层的上一隐藏层,如图1所示,若“本层”为3N1这一隐藏层,则“上一层”为2N1这一隐藏层,以此类推;若“本层”是第一个隐藏层,则“上一层”为输入层,如图1所示,若“本层”为1N1这一隐藏层,则“上一层”为输入层,以此类推;
S2.3、建立解码器的输出层,且所述输出层用于输出包括至少一个用户的码字的解码结果;其中,每个用户对应m=log 2(M)(其中M是每个用户的码字个数)个输出层节点,分别对应解码出来的m个比特向量,由此,所述输出层的节点总数是mJ(其中J是用户数);
S2.4、按照下述公式(2)(也即sigmoid函数)计算至少一个输出层节点的输出概率σ(x):
σ(x)=(1+e -x) -1     (2);
以及按照下述公式(3)计算预测值与真实值之间的不一致程度:
Figure PCTCN2019123144-appb-000018
其中,y=(y 1,…,y 2K) T是上述训练数据集中从物理环境中收集的SCMA信号数据;b=(b 1,…,b mJ) T是与之关联的源码字;W d,b d分别代表整个解码器的权重和偏置集合;d(·;W d,b d)表示解码器的最后输出值,其值是解码出来的比特向量
Figure PCTCN2019123144-appb-000019
此处的
Figure PCTCN2019123144-appb-000020
Figure PCTCN2019123144-appb-000021
也即公式(2)计算出的输出节点的输出概率σ(x),[0,1]为取值范围,π i[·]表示取向量的第i个元素的值;
整体上,如图1所示,本实施例中的SCMA解码器模型为一种基于DNN的模型,包括1个输入层,6个隐藏层以及1个输出层,其中,输入层、隐藏层以及输出层的建立方式参照步骤S2.1-2.4,在此不再赘述;
S3、根据所述训练样本数据集训练上述SCMA解码器模型
本步骤中采用常用的随机梯度下降法对上述SCMA解码器模型进行训练,以搜寻逼近各隐藏层的最优权重和偏置,使得交叉熵损失函数满足下述公式(4)所述条件;
Figure PCTCN2019123144-appb-000022
优选的,本步骤中使用ADAM算法进行SCMA解码器模型训练,由此便于在Tensorflow等人工智能学习系统中直接使用;
同时,训练数据集被分成若干批数据集送入上述模型中进行训练,记每一层的线性输出
Figure PCTCN2019123144-appb-000023
Figure PCTCN2019123144-appb-000024
Figure PCTCN2019123144-appb-000025
Figure PCTCN2019123144-appb-000026
为完整训练数据集在该层的线性输出集合,其中,N b为批数据集大小;
进一步的,为了避免训练过程中可能出现梯度消失和梯度爆炸 问题,本步骤中还引入变量Xavier初始化方法和变量批量归一化方法;具体的,其包括如下步骤:
S3.1变量Xavier初始化,使得每一层(可以是输入层、隐藏层以及输出层中的任意一层)输出的初始状态y l均需要满足以下公式(5)所述条件:
Var[y l]=N l,i·Var[W l]·Var[y l-1]       (5);
其中,Var[·]表示取方差;y l-1是第l-1层输出的初始状态;W l是第l层的权重;N l,i是第l层输入节点总数,且对于前向梯度传递而言,要求N l,i·Var[W l]=1,对于反向梯度传递来说,要求N l,o·Var[W l]=1,N l,o是第l层输出节点总数;优选的,为避免上述两个条件互相抵触,本实施例选取
Figure PCTCN2019123144-appb-000027
对权重进行随机初始化;
S3.2、变量批量归一化:
进一步的,每一层的数据在完成线性输出前均采取以下步骤进行批量归一化:
S3.2.1、对于每个输出节点k∈{1,…,N l,0},分别按照公式(6)、(7)计算均值μ β,k和方差σ β,k 2
Figure PCTCN2019123144-appb-000028
Figure PCTCN2019123144-appb-000029
其中,Z l,k (i)表示
Figure PCTCN2019123144-appb-000030
的向量中的第k个元素,Nb表示完整训练集被分成Nb个批次,i表示第i个批次;
S3.2.2、按照公式(8)进行标准化运算:
Figure PCTCN2019123144-appb-000031
其中,
Figure PCTCN2019123144-appb-000032
ε为常数(为一可训练的常数,其取值由训练结果决定);
S3.2.3、按照公式(9)进行缩放与转换运算,以获得批量归一化结果a l,k i
Figure PCTCN2019123144-appb-000033
其中,
Figure PCTCN2019123144-appb-000034
以及β l,k i是在训练期间与原权重和偏置一起学习得到的缩放转换系数,最终将a l,k i代入公式(1)中进行非线性计算;
S3.3按照步骤S3.1和3.2反复训练SCMA解码器模型,且在训练完成后保存SCMA解码器模型;
S4、将训练后的SCMA解码器模型加载至解码平台,并通过所述SCMA解码器模型对SCMA信号进行解码;其具体包括:
S4.1、将经步骤S3保存的SCMA解码器模型导出成UFF文件格式(universal file format),然后利用TensorRT等AI推理平台将UFF文件格式的SCMA解码器模型优化成.plan文件,再由AIR-T等开发板平台提供的TensorRT GRC模块等导入模块导入,再设置TensorRT GRC模块等导入模块的参数,以将经过步骤S3训练的SCMA解码器模型加载至AIR-T等开发板平台;
S4.2、利用AIR-T等开发板平台提供的无线电接收模块接收SCMA发送机发送的信号,将收到的SCMA信号接入到TensorRT GRC模块等导入模块,进而通过导入模块转至已加载的SCMA解码器模型进行解码处理,编译运行后实时在线解码输出。
实施例二:
本实施例还提供了一种用于实现实施例一中解码方法的SCMA解码设备,如图2所示,其包括:
信号发送机1,其用于将SCMA系统中产生的SCMA信号发送至物理环境中;
信号接收机2,其用于接收物理环境中的SCMA信号,且记录并存储SCMA信号数据;
训练数据生成模块3,其用于将信号发送机1的源码字与SCMA信号数据进行关联,记录并存储关联结果数据,且根据关联结果数据得到不同信噪比条件下的信号数据集,由此得到训练数据集,且将所述训练数据集加载并存储至Numpy数组中;
模型生成模块4,其用于建立基于深度神经网络的SCMA解码器模型;具体的,如图3所示,其包括:输入层建立模块41,其用于接收SCMA信号数据,且具有至少一个资源块;隐藏层建立模块42,其用于进行SCMA信号特征学习与提取,且根据公式(1)完成所述隐藏层的数据输出;输出层建立模块43,其用于根据公式(2)计算至少一个输出层节点的输出概率,且根据所述公式(3)计算预测值与真实值之间的不一致程度,最后输出包括至少一个用户的码字的解码结果;
模型训练成模块5,其用于采用常用的随机梯度下降法对上述SCMA解码器模型进行训练,以搜寻逼近各隐藏层的最优权重和偏置,使得交叉熵损失函数满足公式(4)所述条件,且在训练完成后保存SCMA解码器模型;具体的,如图4所示,所述模型训练成模块5还包括:变量初始化模块51,其用于使得每一层输出的初始状态y l均需要满足公式(5)所述条件;以及变量归一化模块52,其用于将每一层的数据在非线性激活之前进行批量归一化;
模型文件优化模块6,其用于将SCMA解码器模型导出成UFF文件格式,再利用TensorRT等AI推理平台将UFF文件格式的SCMA解码器模型优化成.plan文件;
开发板平台(如AIR-T等)7,如图5所示,其包括导入模块71(如TensorRT GRC模块等)以及无线电接收模块72,用于加载优化后的解码器模型文件,且将接收到的SCMA信号加载至所述解码器模型文件中进行实时在线解码处理,并输出处理结果。
实施例三:
本实施例提供一种无线通信设备,其包括实施例三中所述的解码设备。
传统的SCMA解码方法多采用MPA(Message Pass Algorithm)算法,其复杂度为O(X df),其中,d f为用户过载度,而本发明中的解码器是一种基于DNN的解码器,其复杂度为
Figure PCTCN2019123144-appb-000035
(见表1中的复杂度对比),其中N L为隐藏层层数,,N HN为隐藏层节点数。由计算过程可知,基于传统MPA算法的SCMA解码器的复杂度会随着用户数的增加呈指数级增长,而基于DNN的SCMA解码器的复杂度增 长缓慢。
表1复杂度对比
Figure PCTCN2019123144-appb-000036
例如,对于6个用户共用4个资源块的SCMA特定场景来说,假设1次乘法操作相当于10次加法操作的时间,1次对数或指数操作相当于20次加法操作时间。表2给出在MPA算法迭代5次且信噪比和误码率相同的情况下基于MPA算法的SCMA解码器以及本发明基于DNN的SCMA解码器的复杂度对比情况。由些可见,在这个特定例子中,本发明的基于DNN的SCMA解码器的复杂度较之MPA算法,性能提升了接近50%。
表2复杂度对比举例
Figure PCTCN2019123144-appb-000037
综上所述,本发明提供了一种基于深度神经网络的SCMA解码器,其采用软件无线电平台搭建SCMA系统,同时构建SCMA解码器模型,并将所述SCMA解码器模型部署至AIR-T等开发板上,其操作过程简单、快捷,尤其适用于过载的非正交多址接入无线通信系统,由此可在不增加SCMA解码器复杂度的前提下提高SCMA解码准确率,相对于传统基于MPA算法的SCMA解码器而言,本发明中基于DNN的SCMA解码器在计算复杂度与解码误码率方面性能都有所提升。
需要说明的是,实施例一至三中的技术特征可进行任意组合,且组合而成的技术方案均属于本发明的保护范围。并且,在本文中,诸如术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
尽管已经示出和描述了本发明的实施例,对于本领域的普通技术人员而言,可以理解在不脱离本发明的原理和精神的情况下可以对这些实施例进行多种变化、修改、替换和变型,本发明的范围由所附权利要求及其等同物限定。

Claims (10)

  1. 一种SCMA系统的DNN解码方法,其特征在于,包括如下步骤:
    S1、搭建用于产生SCMA信号的SCMA系统,以及在将SCMA信号发送装置的源码字与SCMA信号接收装置接收到的SCMA信号数据进行关联后获得训练样本数据集;
    S2、建立基于深度神经网络的SCMA解码器模型;
    S3、根据所述训练样本数据集训练上述SCMA解码器模型;
    以及S4、将训练后的SCMA解码器模型加载至解码平台,并通过所述SCMA解码器模型对SCMA信号进行解码。
  2. 如权利要求1所述的解码方法,其特征在于,所述步骤S1包括:
    S1.1、搭建用于产生SCMA信号的SCMA系统,以及搭建SCMA发送机和SCMA接收机,记录并存储SCMA发送机的源码字,并通过SCMA发送机将所述SCMA系统产生的SCMA信号发送至物理环境中;
    S1.2、SCMA接收机接收物理环境中的SCMA信号,且记录并存储SCMA信号数据;
    S1.3、将SCMA发送机的源码字与SCMA接收机存储的SCMA信号数据进行关联,记录并存储关联结果数据;
    S1.4、重复步骤S1.1-1.3,记录并存储每次重复得到的关联结果数据,以此得到不同信噪比条件下的信号数据集;
    以及S1.5、整理不同信噪比条件下的信号数据集中的源码字以及与其关联的SCMA信号数据,由此得到所述训练数据集。
  3. 如权利要求1所述的解码方法,其特征在于,所述步骤S2包括:
    S2.1、建立解码器的输入层,且所述输入层包括至少一个用于接收SCMA信号数据的资源块;
    S2.2、建立解码器的隐藏层,并通过下述公式(1)完成所述隐藏层的数据输出,
    Figure PCTCN2019123144-appb-100001
    其中,y l-1是上一层的数据输出,W l T和b l分别是本层的权重和偏置;
    S2.3、建立解码器的输出层,且所述输出层用于输出包括至少一个用户的码字的解码结果;每个用户对应m=log 2(M)个输出层节点,分别对应解码出来的m个比特向量,所述输出层的节点总数是mJ,其中M是每个用户的码字个数,J是用户数;
    S2.4、按照下述公式(2)计算至少一个输出层节点的输出概率σ(x):
    σ(x)=(1+e -x) -1  (2);
    以及按照下述公式(3)计算预测值与真实值之间的不一致程度:
    Figure PCTCN2019123144-appb-100002
    其中,y=(y 1,…,y 2K) T是上述训练数据集中从物理环境中收集的SCMA信号数据;b=(b 1,…,b mJ) T是与之关联的源码字;W d,b d分别代表整个解码器的权重和偏置集合;d(·;W d,b d)表示解码器的最后输出值,其值是解码出来的比特向量
    Figure PCTCN2019123144-appb-100003
    此处的
    Figure PCTCN2019123144-appb-100004
    Figure PCTCN2019123144-appb-100005
    也即公式(2)计算出的输出节点的输出概率σ(x),[0,1]为取值 范围,π i[·]表示取向量的第i个元素的值。
  4. 如权利要求3所述的解码方法,其特征在于,所述步骤S2.1中,每个资源块设有两个输入节点,分别对应复数信号的实部和虚部。
  5. 如权利要求3所述的解码方法,其特征在于,步骤S3中采用随机梯度下降法对上述SCMA解码器模型进行训练,使得交叉熵损失函数满足下述公式(4)所述条件;
    Figure PCTCN2019123144-appb-100006
  6. 如权利要求5所述的解码方法,其特征在于,步骤S3中还包括:
    S3.1变量Xavier初始化,使得每一层输出的初始状态y l均需要满足以下公式(5)所述条件:
    Var[y l]=N l,i·Var[W l]·Var[y l-1]       (5);
    其中,Var[·]表示取方差;y l-1是第l-1层输出的初始状态;W l是第l层的权重;N l,i是第l层输入节点总数;
    S3.2、将每一层的数据在完成线性输出前进行批量归一化;
    S3.3按照步骤S3.1和3.2反复训练SCMA解码器模型,且在训练完成后保存SCMA解码器模型。
  7. 如权利要求6所述的解码方法,其特征在于,步骤S3.1中,
    Figure PCTCN2019123144-appb-100007
    其中,N l,o是第l层输出节点总数,且对于反向梯度传递而言,要求N l,o·Var[W l]=1。
  8. 如权利要求6所述的解码方法,其特征在于,所述步骤S3.2包括:
    S3.2.1、对于每个输出节点k∈{1,…,N l,0},分别按照公式(6)、(7)计算均值μ β,k和方差σ β,k 2
    Figure PCTCN2019123144-appb-100008
    Figure PCTCN2019123144-appb-100009
    其中,Z l,k (i)表示
    Figure PCTCN2019123144-appb-100010
    的向量中的第k个元素,Nb表示完整训练集被分成Nb个批次,i表示第i个批次;
    S3.2.2、按照公式(8)进行标准化运算:
    Figure PCTCN2019123144-appb-100011
    其中,
    Figure PCTCN2019123144-appb-100012
    i∈{1,…,N b};
    S3.2.3、按照公式(9)进行缩放与转换运算,以获得批量归一化结果a l,k i
    Figure PCTCN2019123144-appb-100013
    其中,
    Figure PCTCN2019123144-appb-100014
    i∈{1,…,N b},γ l,k i以及β l,k i是在训练期间与原权重和偏置一起学习得到的缩放转换系数,最终将a l,k i代入公式(1)中进行非线性计算。
  9. 如权利要求1所述的解码方法,其特征在于,所述步骤S4包括:
    S4.1、将经步骤S3保存的SCMA解码器模型导出并优化,再将其加载至开发板平台;
    S4.2、接收SCMA发送机发送的信号,并将收到的SCMA信号转至 已加载的SCMA解码器模型进行解码处理,编译运行后实时在线解码输出。
  10. 一种用于实现权利要求1-9任一项所述解码方法的解码通信设备,其特征在于,包括:
    训练数据生成模块,其用于将源码字与SCMA信号数据进行关联,且根据关联结果数据得到不同信噪比条件下的信号数据集,由此得到训练数据集;
    模型生成模块,其用于建立基于深度神经网络的SCMA解码器模型;
    模型训练成模块,其用于对上述SCMA解码器模型进行训练,且在训练完成后保存SCMA解码器模型;
    开发板平台,其用于加载优化后的解码器模型文件,且将接收到的SCMA信号加载至所述解码器模型文件中进行实时在线解码处理,并输出处理结果。
PCT/CN2019/123144 2018-12-18 2019-12-05 Scma系统的dnn解码方法及解码通信设备 WO2020125421A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811547598.0 2018-12-18
CN201811547598.0A CN109787715B (zh) 2018-12-18 2018-12-18 Scma系统的dnn解码方法及解码通信设备

Publications (1)

Publication Number Publication Date
WO2020125421A1 true WO2020125421A1 (zh) 2020-06-25

Family

ID=66497057

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/123144 WO2020125421A1 (zh) 2018-12-18 2019-12-05 Scma系统的dnn解码方法及解码通信设备

Country Status (2)

Country Link
CN (1) CN109787715B (zh)
WO (1) WO2020125421A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109787715B (zh) * 2018-12-18 2021-01-19 中国科学院深圳先进技术研究院 Scma系统的dnn解码方法及解码通信设备
CN110636020B (zh) * 2019-08-05 2021-01-19 北京大学 一种自适应通信系统神经网络均衡方法
CN110474716B (zh) * 2019-08-14 2021-09-14 安徽大学 基于降噪自编码器的scma编解码器模型的建立方法
CN113381799B (zh) * 2021-06-08 2022-11-01 哈尔滨工业大学 基于卷积神经网络的低轨星地链路端到端稀疏码多址接入方法
CN113992313B (zh) * 2021-10-25 2023-07-25 安徽大学 一种基于深度学习的均衡网络辅助scma编解码方法
CN114978315B (zh) * 2022-07-28 2022-11-01 南昌大学 基于拟合法的scma辅助可见光通信编解码方法及系统
CN115550934B (zh) * 2022-11-29 2023-04-07 安徽电信规划设计有限责任公司 基于深度学习的混合多址接入异构网络多用户检测方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105356971A (zh) * 2015-10-27 2016-02-24 电子科技大学 一种基于概率计算的scma译码器
US20160254937A1 (en) * 2015-02-27 2016-09-01 Huawei Technologies Co., Ltd Low complexity scma/lds detection systems and methods
CN107508775A (zh) * 2017-07-27 2017-12-22 北京邮电大学 一种稀疏码多址接入系统中的译码方法及装置
CN107743103A (zh) * 2017-10-26 2018-02-27 北京交通大学 基于深度学习的mmtc系统的多节点接入检测和信道估计方法
CN108184128A (zh) * 2018-01-11 2018-06-19 安徽优思天成智能科技有限公司 基于深度神经网络的视频序列丢失帧预测恢复方法
CN109787715A (zh) * 2018-12-18 2019-05-21 中国科学院深圳先进技术研究院 Scma系统的dnn解码方法及解码通信设备

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3342164B1 (en) * 2015-09-03 2020-04-15 MediaTek Inc. Method and apparatus of neural network based processing in video coding
KR102314865B1 (ko) * 2017-03-16 2021-10-19 삼성전자 주식회사 단말이 Temporary User ID(임시 사용자 식별자)를 이용해 5G 네트워크에 등록하는 방안

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160254937A1 (en) * 2015-02-27 2016-09-01 Huawei Technologies Co., Ltd Low complexity scma/lds detection systems and methods
CN105356971A (zh) * 2015-10-27 2016-02-24 电子科技大学 一种基于概率计算的scma译码器
CN107508775A (zh) * 2017-07-27 2017-12-22 北京邮电大学 一种稀疏码多址接入系统中的译码方法及装置
CN107743103A (zh) * 2017-10-26 2018-02-27 北京交通大学 基于深度学习的mmtc系统的多节点接入检测和信道估计方法
CN108184128A (zh) * 2018-01-11 2018-06-19 安徽优思天成智能科技有限公司 基于深度神经网络的视频序列丢失帧预测恢复方法
CN109787715A (zh) * 2018-12-18 2019-05-21 中国科学院深圳先进技术研究院 Scma系统的dnn解码方法及解码通信设备

Also Published As

Publication number Publication date
CN109787715B (zh) 2021-01-19
CN109787715A (zh) 2019-05-21

Similar Documents

Publication Publication Date Title
WO2020125421A1 (zh) Scma系统的dnn解码方法及解码通信设备
Ahn et al. Wireless federated distillation for distributed edge learning with heterogeneous data
KR101618023B1 (ko) 희소 코드 다중 액세스를 위한 시스템 및 방법
CN114514519A (zh) 使用异构模型类型和架构的联合学习
Yang et al. Deep multimodal learning: Merging sensory data for massive MIMO channel prediction
CN112491442B (zh) 一种自干扰消除方法及装置
CN111478868B (zh) 一种otfs系统的信号检测方法及装置
CN107425889B (zh) 一种5g通信系统接收端基带信号联合处理方法
Sery et al. COTAF: Convergent over-the-air federated learning
CN113965233A (zh) 一种基于深度学习的多用户宽带毫米波通信资源分配方法及系统
Xie et al. Massive unsourced random access for massive MIMO correlated channels
Cheng et al. Deep neural network aided low-complexity MPA receivers for uplink SCMA systems
Ya et al. Modulation recognition of digital signal based on deep auto-ancoder network
Fan et al. Intelligent communication: Application of deep learning at the physical layer of communication
Yang et al. Efficient hardware architecture of deterministic MPA decoder for SCMA
CN109586848A (zh) Scma系统中一种消息传递解码算法
Shevchuk Theoretical and algorithmic foundations of improving the efficiency of packet data transmission in high-speed and secure radio networks
CN115052288A (zh) 一种基于深度学习的多用户物理层认证方法
CN112261685B (zh) 一种基于卷积神经网络的mimo信号检测方法
Lin et al. Communication-Efficient Joint Signal Compression and Activity Detection in Cell-Free Massive MIMO
Liu et al. Flexible beamforming of dynamic MIMO networks through fully convolutional model
CN108737298B (zh) 一种基于图像处理的scma盲检测方法
Wong et al. Distributed automatic modulation classification with compressed data
Wang et al. Deep learning-based automatic modulation recognition algorithm in Internet of Things
Wang et al. Sparse power control for downlink cell-free massive MIMO systems with limited backhaul capacities

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19899969

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1255 DATED 10.08.2021.)

122 Ep: pct application non-entry in european phase

Ref document number: 19899969

Country of ref document: EP

Kind code of ref document: A1