CN116527180A - SCMA method based on CWGAN-GP satellite-ground link channel modeling - Google Patents

SCMA method based on CWGAN-GP satellite-ground link channel modeling Download PDF

Info

Publication number
CN116527180A
CN116527180A CN202310535528.8A CN202310535528A CN116527180A CN 116527180 A CN116527180 A CN 116527180A CN 202310535528 A CN202310535528 A CN 202310535528A CN 116527180 A CN116527180 A CN 116527180A
Authority
CN
China
Prior art keywords
cwgan
satellite
link channel
ground link
scma
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
Application number
CN202310535528.8A
Other languages
Chinese (zh)
Inventor
李东博
孙宇辰
刘劼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
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 Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202310535528.8A priority Critical patent/CN116527180A/en
Publication of CN116527180A publication Critical patent/CN116527180A/en
Pending legal-status Critical Current

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • H04B17/3912—Simulation models, e.g. distribution of spectral power density or received signal strength indicator [RSSI] for a given geographic region
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/14—Relay systems
    • H04B7/15—Active relay systems
    • H04B7/185—Space-based or airborne stations; Stations for satellite systems
    • H04B7/1851—Systems using a satellite or space-based relay
    • H04B7/18513—Transmission in a satellite or space-based system
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/14—Relay systems
    • H04B7/15—Active relay systems
    • H04B7/185—Space-based or airborne stations; Stations for satellite systems
    • H04B7/1851—Systems using a satellite or space-based relay
    • H04B7/18519—Operations control, administration or maintenance
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/14—Relay systems
    • H04B7/15—Active relay systems
    • H04B7/204—Multiple access
    • H04B7/216—Code division or spread-spectrum multiple access [CDMA, SSMA]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Astronomy & Astrophysics (AREA)
  • Electromagnetism (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Radio Relay Systems (AREA)

Abstract

The SCMA method based on CWGAN-GP satellite-ground link channel modeling solves the problem of reduced information transmission reliability in the prior art, and belongs to the technical field of information and communication. The invention comprises the following steps: constructing a sparse code multiple access model based on a convolutional neural network, wherein the sparse code multiple access model comprises an encoder, a decoder and a satellite-to-ground link channel; the star-to-ground link channel model is realized by adopting a gradient penalty-based condition Wasserstein generation countermeasure network, and comprises a generator and a discriminator which are realized by adopting a plurality of convolutional neural network units; determining parameters of a generator and a discriminator by adopting end-to-end training, completing satellite-to-ground link channel modeling, and performing sparse code multiple access; the invention adopts Wasserstein distance to meet Lipschitz constraint conditions, takes the information coded by an encoder as condition information, adopts gradient penalty to solve the problem that partial data cannot be converged due to forced meeting Lipschitz constraint by weight clipping, and constructs approximate accurate condition channel distribution.

Description

SCMA method based on CWGAN-GP satellite-ground link channel modeling
Technical Field
The invention relates to an SCMA method based on CWGAN-GP satellite-ground link channel modeling, and belongs to the technical field of information and communication.
Background
Under the high dynamic low orbit satellite network topology, the motion characteristics of the large-scale internet of things terminal node are generally random, so that the link dynamic change is caused, the channel state is directly changed, good channel state information is difficult to obtain, and the information transmission reliability is directly reduced. Under the satellite-ground link of the low-orbit satellite network, accurate channel state information can help to better estimate channel quality, improve the accuracy of data transmission, better manage network resources and improve network efficiency. At present, a traditional mathematical channel modeling method is adopted for satellite-ground link channel modeling, but an accurate channel state cannot be truly reflected in many scenes. Therefore, in order to meet the requirements of large-scale internet of things terminal access and future intelligent end-to-end network systems, it is highly desirable to propose a precise channel modeling method, which can construct approximately precise conditional channel distribution even in a dynamic network topology environment, thereby implementing intelligent channel modeling and reducing system complexity.
In the future satellite internet of things multi-user actual communication scenario, a communication system based on traditional Sparse Code Multiple Access (SCMA) is limited in many aspects, for example, due to the fact that a multidimensional codebook design needs to use an information transfer algorithm (MPA) for decoding, high computational overhead is caused at a decoding end, and due to the difference of the number of resources in different communication environments, it is necessary to manually construct a codebook for all possible actual communication scenarios, meanwhile, channel modeling of traditional sparse code multiple access is expressed as an embedded assumed mathematical model, an actual transmission scheme cannot be accurately reflected, and system performance is lost.
Disclosure of Invention
Aiming at the problems that under the high-dynamic low-orbit satellite network topology, the motion characteristics of large-scale Internet of things terminal nodes are generally random, link dynamic changes are caused, channel states are directly changed, good channel state information is difficult to obtain, and the reliability of information transmission is directly reduced, the invention provides an SCMA method based on CWGAN-GP satellite-to-ground link channel modeling.
The invention discloses an SCMA method based on CWGAN-GP satellite-ground link channel modeling, which comprises the following steps:
s1, constructing a convolutional neural network sparse code multiple access SCMA model, wherein the model comprises an encoder of the sparse code multiple access SCMA, a satellite-ground link channel and a decoder of the sparse code multiple access SCMA;
the satellite-ground link channel is realized by adopting a generating countermeasure network CWGAN-GP, and comprises a generator G (z|m; theta) realized by adopting a plurality of convolution neural network units G ) And a discriminator D (x|m; θ D )；
G(z|m；θ G ) Representing the random noise distribution P to be followed z The mid-sampled noise samples z are transformed to conform to the profile P g The noise samples z conform to condition information m including the corresponding pilot symbols y p Is a signal received by the base station; θ G Parameters representing the arbiter;
D(x|m；θ D ) Is the true data distribution P r Real sample x and generator G (z|m; θ) G ) The obtained analog sample, the real sample x is the output of the encoder; the real sample x and the simulation sample accord with the condition information m, theta D Parameters representing the arbiter; d (x|m; θ) D ) The output of (2) is a probability;
s2, training a convolutional neural network sparse code multiple access SCMA model, determining parameters of a generator and a discriminator, completing satellite-ground link channel modeling, and performing sparse code multiple access SCMA;
in the training phase, the loss function of the counter network CWGAN-GP is generated as follows:
wherein P is s Represented in the true sample distribution P r And generating a sample distribution P g A sample space formed by uniformly sampling straight lines between the point pairs;
D(G(z|m；θ D ) Representing the coincidence distribution P g Is input to a discriminator D (x|m; θ) D ) After the middle, the obtained probability; I.I 2 Representing the two norms of the two-way model,representation->For->Is a partial derivative of (2);
representing random noise distribution P z The lower noise sample z is equal to D (G (z|m; θ) D ) A) the desire to do so;
representing a true data distribution P r The lower real sample x vs D (x|m; θ) D ) Is not limited to the desired one;
represented in sample space P s Downsampling sample +.>For->Is not limited to the desired one; kappa represents a penalty coefficient.
Preferably, the encoder of the sparse code multiple access SCMA is composed of a basic convolutional neural network unit with a plurality of hidden layers;
the encoder adopts a convolutional neural network unit to realize the mapping from the data streams r of J users to constellation planes of K resource blocks, K is smaller than J, and a real sample with the data length of N is obtained after encoding
Preferably, the sample is sampledThe method comprises the following steps:
wherein ε is uniformly distributed according to [0,1 ].
Preferably, the decoder of the sparse code multiple access SCMA is composed of a basic convolutional neural network unit with a plurality of hidden layers; the decoder adopts a convolutional neural network unit to copy the data streams of J users according to the received signal y-intermediate Xi Hui received from the satellite-ground link channelAnd output.
Preferably, the parameters of the decoder are updated using a binary cross entropy loss function:
wherein,,an mth information bit representing a jth user data stream input by the encoder,/-, and a bit representing a mth user data stream input by the encoder>An mth information bit representing a jth user data stream output by the decoder.
Preferably, S2 includes:
s21, determining a training data set comprising data flows r of J users, received signals y correspondingly received by a decoder and received pilot data y p ；
S22, sampling data flows r of J users and corresponding received signals y, and updating parameters of a decoder by utilizing the binary cross entropy loss function until the set iteration times are reached;
s23, data flows r, received signals y and pilot data sets y of J users p Sampling noise, forming a sparse code multiple access SCMA end-to-end system by using an encoder, a generator and a decoder, calculating a loss function of the sparse code multiple access SCMA end-to-end system, and updating parameters of the encoder, the generator and the decoder based on SGD until the set iteration times are reached;
s24, sampling the data flows r of J users and the instantaneous state information h of the satellite-ground link channel, and receiving the received signal y and the pilot frequency data set y therein p Sampling the noise sample z, sampling the real sample distribution P r And generating a sample distribution P g A sample space P formed by uniformly sampling a straight line between pairs of points s Sampling, calculating and generating a loss function of the counter network CWGAN-GP, and updating the parameter theta of the generator G And parameter θ of the arbiter D Until the set iteration times are reached.
Preferably, the output of the ith neuron in the first full connection layer of the convolutional neural network unit is:
wherein, sigma (·) is the activation function,indicating that it is connected to the first degreel-1) weights of the jth neuron in the fully connected layer and the ith neuron in the first fully connected layer, Z (l-1) [j]Representing the output of the jth neuron in the first-1 fully-connected layer, b i (l) Representing the bias vector of the ith neuron in the first fully connected layer.
Preferably, the output of the ith neuron in the first convolutional layer of the convolutional neural network unit is:
wherein, sigma (·) is the activation function,is a convolution coefficient, +.>Representing the bias vector of the ith neuron in the first convolutional layer.
The invention has the beneficial effects that under the satellite-ground link of the low-orbit satellite network, accurate channel state information can help to better estimate the channel quality, improve the accuracy of data transmission, better manage network resources and improve the network efficiency. At present, a traditional mathematical channel modeling method is adopted for satellite-ground link channel modeling, but an accurate channel state cannot be truly reflected in many scenes. Therefore, in order to meet the requirements of large-scale internet of things terminal access and future intelligent end-to-end network systems, it is highly desirable to propose a precise channel modeling method, which can construct approximately precise conditional channel distribution even in a dynamic network topology environment, thereby implementing intelligent channel modeling and reducing system complexity. The method is based on the self-encoder to construct an end-to-end intelligent communication system, and replaces the traditional modularized transmission model. The method comprises the steps of constructing a sparse code multiple access encoder and decoder by adopting a convolutional neural network, providing a condition Wasserstein generation countermeasure network (CWGAN-GP) based on gradient penalty optimization to model a satellite-ground link channel of a low-orbit satellite network, adopting a Wasserstein distance to meet Lipschitz constraint conditions, taking information coded by the convolutional neural network encoder as condition information, adopting gradient penalty solution weight clipping to forcibly meet Lipschitz constraint so as to cause the problem that partial data cannot be converged, and realizing construction of approximate accurate condition channel distribution based on the CWGAN-GP. The back propagation of the gradient is realized through the established channel model, meanwhile, the updating of the parameters of the encoder and the decoder is realized through the iterative training of the network, the accurate intelligent channel modeling under the low orbit satellite dynamic network is realized, and the complexity of the system is reduced.
Drawings
FIG. 1 is a block diagram of a CNN-SCMA for generating a model of an antagonistic network satellite-to-ground link channel based on Wasserstein optimization of gradient penalties;
FIG. 2 is a block diagram of smart channel modeling for generating an countermeasure network based on gradient penalty Wasserstein;
FIG. 3 is a graph showing BER performance under a satellite-to-ground link of a low-orbit satellite network, with SNR on the abscissa;
fig. 4 shows the block error rate BLER performance under a low-orbit satellite network satellite-ground link, with the abscissa representing the signal-to-noise ratio SNR.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.
In order to solve the problems that under the high-dynamic low-orbit satellite network topology, the motion characteristics of the large-scale Internet of things terminal node are generally random, the link dynamic change is caused, the channel state is directly changed, good channel state information is difficult to obtain, and the reliability of information transmission is directly reduced. The SCMA method based on CWGAN-GP satellite-ground link channel modeling in the embodiment comprises the following steps:
step 1, constructing a convolutional neural network sparse code multiple access model, wherein the convolutional neural network sparse code multiple access model comprises a sparse code multiple access encoder, a satellite-to-ground link channel and a sparse code multiple access decoder;
the satellite-ground link channel comprises a generator G (z|m; θ) implemented with a generation counter network CWGAN-GP, including a plurality of convolutional neural network elements G ) And a discriminator D (x|m; θ D )；
Generator G (z; θ) for constructing CWGAN-GP G )，θ G Is CWGAN-GP generator G (z; θ) G ) Is set, the generator G (z; θ G ) Mainly will be distributed from P z The mid-sampled noise samples z are transformed to conform to the profile P g Is a simulated sample of (a);
construction of a discriminator D (x; θ) of CWGAN-GP D )，θ D D (x; θ) being a CWGAN-GP arbiter D ) X is the sample, and the arbiter D (x; θ D ) The input of (a) is from the true data distribution P r True samples of the sum generator G (z; θ) G ) The generated analog sample, and the output of the arbiter D is a probability;
if the sample x is identified as possibly being distributed from the real data P r Extracted from the above, the decision device D (x; θ) D ) The probability obtained is close to 1; otherwise, the arbiter network D (x; θ D ) A probability of approaching 0 will be output. During training, generator G (z; θ) G ) Will attempt to produce a channel similar to the real channel P r Output samples, and a discriminator D (x; θ) D ) Will attempt to distinguish between the channels from the true channel P r Data from generator G (z; θ) G ) Is a data of (a) a data of (b).
Corresponding to pilot symbol y p Is added as part of the condition information m to give the input x and the received pilot data y p In the case of (2), the output samples follow the y distribution. And a generator of CWGAN-GP channelG(z|m；θ G ) And a discriminator D (x|m; θ D ) The extra information m is conditioned.
WGAN is used instead of GAN to improve system training stability. Unlike Jensen-Shannon divergence employed by GAN, WGAN employs waserstein distance to satisfy Lipschitz constraints. A Gradient Penalty (GP) based CWGAN network is proposed, wherein the GP is added to an optimization function by limiting the Gradient norm of the network output relative to its input, and the Gradient norm is limited to 1 to ensure 1-Lipschitz continuity.
The present embodiment connects the encoder and decoder through a channel modeled based on CWGAN-GP. Since the channel output y is determined by the condition distribution p (y|x) given the input x, an approximately accurate condition channel distribution can be constructed by CWGAN-GP with x as condition information. The CWGAN-GP channel modeling can realize the gradient transfer from the decoder to the encoder, thereby realizing the end-to-end training of CNN-based sparse code multiple access. A block diagram of the smart channel modeling for generating the challenge network CWGAN-GP based on the gradient penalty wasperstein is shown in fig. 2.
CWGAN-GP is mainly derived from GAN, and GAN is mainly composed of generator G and discriminator D, and its parameters are respectively set to θ G And theta D . In GAN, generator G (z; θ) G ) Mainly will be distributed from P z The mid-sampled noise samples z are transformed to conform to the profile P g Is a sample of the analog sample of (a). Distinguishing device D (x; θ) D ) The input of (a) is from the true data distribution P r True samples of the sum generator G (z; θ) G ) The analog sample is generated and the output of the arbiter D is a probability. If the sample x is identified as possibly being distributed from the real data P r Extracted from the above, the decision device D (x; θ) D ) The probability obtained is close to 1; otherwise, the arbiter network D (x; θ D ) A probability of approaching 0 will be output. During training, generator G (z; θ) G ) Will attempt to produce a channel similar to the real channel P r Output samples, and a discriminator D (x; θ) D ) Will attempt to distinguish between the channels from the true channel P r Data from generator G (z; θ) G ) Is a data of (a) a data of (b). Raw materialsThe optimized objective function of the constructor and arbiter can be expressed as
Y received at the decoder can be expressed as:
y＝hx+n
where H is the instantaneous CSI of the channel, obtained by sampling the channel set H, and n is the noise vector of the channel. Acquisition of instantaneous CSI is critical to the decoding of the system, so in our approach, will correspond to pilot symbols y p Is added as part of the condition information m to give the input x and the received pilot data y p In the case of (2), the output samples follow the y distribution. With the pilot information, the receiver can obtain information about the specific channel implementation, which is very helpful for recovering the transmitted data. Meanwhile, by utilizing pilot frequency data, the generation of channel output distribution under the current channel based on the learning of the CWGAN-GP channel generator can be realized.
Corresponding to pilot symbol y p Is added as part of the condition information m to give the input x and the received pilot data y p In the case of (2), the output samples follow the y distribution. And generator G (z|m; θ) of CWGAN-GP channel G ) And a discriminator D (x|m; θ D ) The extra information m is conditioned.
Meanwhile, the WGAN is adopted to replace GAN to improve the training stability of the system. Unlike Jensen-Shannon divergence employed by GAN, WGAN employs waserstein distance to satisfy Lipschitz constraints. The objective function of CWGAN is
However, since WGAN uses Weight Clipping (Weight Clipping) to force the Lipschitz constraint to be satisfied, CWGAN still creates a sample non-convergence situation in some cases. Therefore, a CWGAN network based on Gradient Penalty (GP) is proposed, by limiting the Gradient norms of the network output with respect to its input, adding GP to the optimization function, and limiting the Gradient norms to 1, to ensure 1-Lipschitz continuity. The gradient penalty term is expressed as
Wherein I II 2 Representing the two norms of the two-way model,representation->For->Partial derivative of>Is distribution P s Is obtained by sampling in a sample space of (a) its distribution P s Is defined as being distributed over the true samples P r And generating a sample distribution P g A sample space formed by uniform sampling of the straight line between the pairs of points>Can be expressed as
Wherein ε obeys a uniform distribution of [0,1 ].
Thus, the loss function generated against the network in the CWGAN-GP architecture proposed by the embodiment is expressed as
Wherein P is s Expressed in realitySample distribution P r And generating a sample distribution P g A sample space formed by uniformly sampling straight lines between the point pairs;
D(G(z|m；θ D ) Representing the coincidence distribution P g Is input to a discriminator D (x|m; θ) D ) After the middle, the obtained probability; I.I 2 Representing the two norms of the two-way model,representation->For->Is a partial derivative of (2);
representing random noise distribution P z The lower noise sample z is equal to D (G (z|m; θ) D ) A) the desire to do so;
representing a true data distribution P r The lower real sample x vs D (x|m; θ) D ) Is not limited to the desired one;
represented in sample space P s Downsampling sample +.>For->Is not limited to the desired one; kappa represents a penalty coefficient. .
Step 2, training a convolutional neural network sparse code multiple access model, determining parameters of a generator and a discriminator, and completing satellite-ground link channel modeling;
aiming at the problem that good channel state information is difficult to obtain due to random channel change caused by the high dynamic characteristics of nodes in the low-orbit satellite Internet of things, the embodiment provides a method for generating an antagonistic network channel modeling based on gradient penalty Wasserstein under a low-orbit satellite-ground link. The method proposes generating an antagonism network based on a gradient penalty wasperstein for modeling a channel distribution to represent the channel effects of a low-orbit satellite network satellite-ground link.
In a preferred embodiment, the encoder of this embodiment is constructed of a basic convolutional neural network element having a plurality of hidden layers;
in this embodiment, at the transmitting end of the system, a CNN-based encoder is given according to the AE structure, so as to implement encoding of the sparse code multiple access encoder. The mapping from the data streams of J users to the constellation planes of K resource blocks is realized by adopting CNN, and K is less than J. Encoder f of a system e (. Cndot.) and decoder f d (. Cnc.) is composed of a basic CNN unit with multiple hidden layers, thus, encoder f e (·) can be seen as a code word generator for sparse code multiple access. Each hidden layer of the basic CNN unit is composed of a weight matrix W l Bias vector b l And an activation function phi l Composition, where l represents the index of the hidden layer.
The data input of CNN-based encoder is a data stream r of information bits of length M, i.er j Representing the data stream of the jth user. Definition of theta e Weights W for CNN-based encoders l And offset vector b l ，/>Is defined as a constellation mapping of a jth user data stream to a kth resource block based on CNN, and the data of the kth resource block encoded by the CNN-based encoder can be expressed as
The data streams of J users are coded by a CNN-based coder and then converted into data with the data length of NAnd sent to the channel.
In the preferred embodiment the decoder is comprised of a basic convolutional neural network element having a plurality of hidden layers; at the receiving end, the received signal of the kth resource block of the CNN-based decoder may be written as follows:
the CNN-based decoder learns to recover the original information from the signal y received from the channel. Definition of theta d Weights W for CNN-based decoders l And offset vector b l The J users' data streams to be restored, which are output by it, can be expressed as:
at the receiving end, calculating the data flows r of the original J users and the restored data flows of the J users by adopting a binary cross entropy loss functionDistance between them. The binary cross entropy loss function can be expressed as:
wherein the method comprises the steps ofAn mth information bit representing a jth user data stream input by the sparse code multiple access encoder,an mth information bit representing a jth user data stream output by the sparse code multiple access decoder.
The embodiment mainly adopts a convolutional neural network in encoder, decoder and CWGAN-GP channel modeling-based generator and discriminator, and comprises a convolutional layer, a full-connection layer and an activation function.
Will Z (l-1) [i]And Z (l) [i]Respectively denoted as input and output of the ith neuron in the first layer of CNN. For the fully connected layer, the output of the ith neuron in the first layer is
Where σ (·) is the activation function, in the present system, a rectifying linear unit (ReLU) is used as the activation function.Representing the weight of the j-th neuron connected in the (l-1) -th layer and the i-th neuron in the first layer,/->Representing the bias vector of the ith neuron in the first layer.
On the other hand, for the convolutional layer, the output of the ith neuron in the first layer is
Wherein the method comprises the steps ofIs a convolution coefficient. The connections and weights between adjacent layers of the convolution layer are much less than for the fully connected layer, which reduces the complexity and significantly increases the convergence speed of the training. The weight of the entire CNN is updated with SGD, where the calculated loss gradient will be derived fromThe output layer propagates back to the input layer.
In this embodiment, training is performed on a CNN-SCMA based on CWGAN-GP channel modeling, and the training process includes:
to obtain training data set, first, data flows r of J users are randomly generated, the received sparse code multiple access signal y and the received pilot data y are mapped to K resource blocks and then to channels through a CNN-based encoder p And the originally transmitted data is collected as training data. The training process based on CWGAN-GP channel modeling is shown in detail in algorithm 1.
The main goal of the decoder is to train a CNN-based decoder to recover the input signal r of the sparse code multiple access encoder. The sparse code multiple access data received by the decoder comprises pilot symbols y p And receiving data y, and the output is an estimate obtained by a CNN-based decoder
The CNN-based decoder takes the received sparse code multiple access data as input and recovers the transmitted data in an end-to-end manner. By calculating the loss function, the receiver is trained and the gradient of the loss is obtained. For time-varying channels, comprising two training phases, in an off-line training phase, by directly multiplexing the received sparse code multiple access signal y with the received pilot data y p Put together as input, the model is trained, and the data is generated from various information sequences and under various channel conditions with specific statistical properties.
In the online deployment stage, output generated based on CNN modelThe transmitted data can be recovered without explicitly estimating the wireless channel. In training the encoder, the training of the CNN-based encoder will be similar to the training of the CNN-based decoder, with the generator of the CWGAN-GP channel as a proxy channel.
The end-to-end cross entropy loss is calculated at the decoder and the gradient propagates back to the transmitter over the CWGAN-GP channel. The weights of the CNN-based encoder will be updated based on the SGD, while the weights of the CWGAN-GP channel and the receiver remain fixed.
In each iteration, the generator and the arbiter are iteratively trained. The parameters of one model will be fixed while the other model is being trained. By the learned transmitter, the real data can be obtained using the encoded signal from the transmitter via the real channel, while the analog data is obtained from the encoded data via the channel generator. Parameters of the generator and the arbiter are updated according to the loss function.
Performance simulation analysis:
in the embodiment, super-parameters, network model parameter design and simulation demonstration are carried out on CNN-SCMA based on CWGAN-GP channels, and simulation results are given. BER and BLER performance of the system were demonstrated and analyzed under different channel conditions for the AWGN, rayleigh and low-orbit satellite network satellite-ground link Lutz, respectively.
The ability to model the channel effect in a data driven manner based on the CWGAN-GP can be seen in the demonstration process, and it is verified that the end-to-end CNN-SCMA system based on the CWGAN-GP channel can obtain similar or better results than the sparse code multiple access based on the self-encoder under the condition that the channel information is not known when the transmitter and the receiver are trained and optimized. In the face of high complexity multi-user systems, the CWGAN-GP channel can still accurately learn various different types of random channel models from measurements without introducing many assumptions about the effects that occur or simplifying into parametric models.
The super parameters of CNN-SCMA based on CWGAN-GP channel modeling are shown in table 1, the system mainly adopts SGD to update the weights of a CNN-based encoder and a CNN-based decoder, adopts Adam to update the weights of a generator and a discriminator of a channel, and adopts ReLU as an activation function.
TABLE 1 super parameters of CNN-SCMA based on CWGAN-GP channel
In the system, a convolutional neural network is mainly adopted in an encoder, a decoder and a generator and a discriminator based on CWGAN-GP channel modeling, and comprises a convolutional layer, a full-connection layer and an activation function, wherein model parameters are shown in a table 2.
TABLE2 model parameters Table2 Model parameters of CNN-SCMA based on CWGAN-GP channel for CNN-SCMA based on CWGAN-GP channel
/>
Figures 3 and 4 show BER and BLER performance, respectively, for the satellite-to-ground link Lutz channel. The BER performance of CNN-SCMA based on CWGAN-GP channel is slightly higher than end-to-end sparse code multiple access based on deep learning, and the BLER performance of end-to-end sparse code multiple access based on deep learning is slightly higher than CNN-SCMA based on CWGAN-GP channel, and these simulations prove that the effectiveness of channel modeling based on CWGAN-GP is enough to prove under end-to-end CNN-SCMA.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other described embodiments.

Claims (10)

1. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling is characterized by comprising the following steps:
s1, constructing a convolutional neural network sparse code multiple access SCMA model, wherein the model comprises an encoder of the sparse code multiple access SCMA, a satellite-ground link channel and a decoder of the sparse code multiple access SCMA;
the satellite-ground link channel is realized by adopting a generating countermeasure network CWGAN-GP, and comprises a generator G (z|m; theta) realized by adopting a plurality of convolution neural network units G ) And a discriminator D (x|m; θ D )；
G(z|m；θ G ) Representing the random noise distribution P to be followed z The mid-sampled noise samples z are transformed to conform to the profile P g The noise samples z conform to condition information m including the corresponding pilot symbols y p Is a signal received by the base station; θ G Parameters representing the arbiter;
D(x|m；θ D ) Is the true data distribution P r Real sample x and generator G (z|m; θ) G ) The obtained analog sample, the real sample x is the output of the encoder; the real sample x and the simulation sample accord with the condition information m, theta D Parameters representing the arbiter; d (x|m; θ) D ) The output of (2) is a probability;
s2, training a convolutional neural network sparse code multiple access SCMA model, determining parameters of a generator and a discriminator, completing satellite-ground link channel modeling, and performing sparse code multiple access SCMA;
in the training phase, the loss function of the counter network CWGAN-GP is generated as follows:
wherein P is s Represented in the true sample distribution P r And generating a sample distribution P g A sample space formed by uniformly sampling straight lines between the point pairs;
D(G(z|m；θ D ) Representing the coincidence distribution P g Is input to a discriminator D (x|m; θ) D ) After the middle, the obtained probability; I.I 2 Representing the two norms of the two-way model,representation->For->Is a partial derivative of (2);
representing random noise distribution P z The lower noise sample z is equal to D (G (z|m; θ) D ) A) the desire to do so;representing a true data distribution P r The lower real sample x vs D (x|m; θ) D ) Is not limited to the desired one;
represented in sample space P s Downsampling sample +.>For->Is not limited to the desired one; kappa represents a penalty coefficient.
2. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 1, wherein the encoder of sparse code multiple access SCMA is comprised of a basic convolutional neural network element having multiple hidden layers;
the encoder adopts a convolutional neural network unit to realize the mapping from the data streams r of J users to constellation planes of K resource blocks, K is smaller than J, and a real sample with the data length of N is obtained after encoding
3. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 1, wherein samples are sampledThe method comprises the following steps:
wherein ε is uniformly distributed according to [0,1 ].
4. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 3, wherein the decoder of the sparse code multiple access SCMA is composed of a basic convolutional neural network unit having a plurality of hidden layers; the decoder adopts a convolutional neural network unit to copy the data streams of J users according to the received signal y-intermediate Xi Hui received from the satellite-ground link channelAnd output.
5. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 4, wherein parameters of the decoder are updated with binary cross entropy loss functions:
wherein,,an mth information bit representing a jth user data stream input by the encoder,/-, and a bit representing a mth user data stream input by the encoder>An mth information bit representing a jth user data stream output by the decoder.
6. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 5, wherein the S2 comprises:
s21, determining a training data set comprising data flows r of J users, received signals y correspondingly received by a decoder and received pilot data y p ；
S22, sampling data flows r of J users and corresponding received signals y, and updating parameters of a decoder by utilizing the binary cross entropy loss function until the set iteration times are reached;
s23, data flows r, received signals y and pilot data sets y of J users p Sampling noise, forming a sparse code multiple access SCMA end-to-end system by using an encoder, a generator and a decoder, calculating a loss function of the sparse code multiple access SCMA end-to-end system, and updating parameters of the encoder, the generator and the decoder based on SGD until the set iteration times are reached;
s24, sampling the data flows r of J users and the instantaneous state information h of the satellite-ground link channel, and receiving the received signal y and the pilot frequency data set y therein p Sampling the noise sample z, sampling the real sample distribution P r And generating a sample distribution P g A sample space P formed by uniformly sampling a straight line between pairs of points s Sampling, calculating and generating a loss function of the counter network CWGAN-GP, and updating the parameter theta of the generator G And parameter θ of the arbiter D Until the set iteration times are reached.
7. The SCMA method based on CWGAN-GP satellite-to-ground link channel modeling of claim 4, wherein the output of the ith neuron in the first fully-connected layer of the convolutional neural network element is:
wherein, sigma (·) is the activation function,representing the weight of the jth neuron connected in the (l-1) th fully connected layer and the ith neuron in the first fully connected layer, Z (l-1) [j]Representing the output of the j-th neuron in the 1-th fully-connected layer,/th neuron>Representing the bias vector of the ith neuron in the first fully connected layer.
8. The SCMA method based on CWGAN-GP constellation-earth link channel modeling according to claim 4, wherein the output of the ith neuron in the first convolutional layer of the convolutional neural network unit is:
wherein, sigma (·) is the activation function,is a convolution coefficient, +.>Representing the bias vector of the ith neuron in the first convolutional layer.
9. A computer readable storage device storing a computer program, characterized in that the computer program when executed implements the SCMA method based on CWGAN-GP star-to-ground link channel modeling according to any of claims 1 to 8.
10. A wasperstein-based star-to-ground link channel modeling apparatus for generating a challenge network, comprising a storage device, a processor and a computer program stored in the storage device and executable on the processor, characterized in that the processor executes the computer program to implement the SCMA method of CWGAN-GP star-to-ground link channel modeling according to any of claims 1 to 8.
CN202310535528.8A 2023-05-12 2023-05-12 SCMA method based on CWGAN-GP satellite-ground link channel modeling Pending CN116527180A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310535528.8A CN116527180A (en) 2023-05-12 2023-05-12 SCMA method based on CWGAN-GP satellite-ground link channel modeling

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310535528.8A CN116527180A (en) 2023-05-12 2023-05-12 SCMA method based on CWGAN-GP satellite-ground link channel modeling

Publications (1)

Publication Number Publication Date
CN116527180A true CN116527180A (en) 2023-08-01

Family

ID=87408068

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310535528.8A Pending CN116527180A (en) 2023-05-12 2023-05-12 SCMA method based on CWGAN-GP satellite-ground link channel modeling

Country Status (1)

Country Link
CN (1) CN116527180A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117914656A (en) * 2024-03-13 2024-04-19 北京航空航天大学 End-to-end communication system design method based on neural network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117914656A (en) * 2024-03-13 2024-04-19 北京航空航天大学 End-to-end communication system design method based on neural network
CN117914656B (en) * 2024-03-13 2024-05-10 北京航空航天大学 End-to-end communication system design method based on neural network

Similar Documents

Publication Publication Date Title
CN110474716B (en) Method for establishing SCMA codec model based on noise reduction self-encoder
CN108566257B (en) Signal recovery method based on back propagation neural network
Ye et al. Deep-learning-enhanced NOMA transceiver design for massive MTC: Challenges, state of the art, and future directions
CN113381828B (en) Sparse code multiple access random channel modeling method based on condition generation countermeasure network
CN116527180A (en) SCMA method based on CWGAN-GP satellite-ground link channel modeling
CN111711455B (en) Polarization code BP decoding method based on neural network
CN109981224A (en) A kind of deep space communication channel decoding system and method
CN114362859A (en) Adaptive channel modeling method and system for enhanced conditional generation countermeasure network
CN110278570B (en) Wireless communication system based on artificial intelligence
Bo et al. Learning based joint coding-modulation for digital semantic communication systems
CN114513394B (en) Signal modulation format identification method, system and device based on attention mechanism diagram neural network and storage medium
CN113381799A (en) Low-orbit satellite-ground link end-to-end sparse code multiple access method based on convolutional neural network
CN110474798A (en) A method of wireless communication future signal is predicted using echo state network
Tabi et al. Hybrid Quantum-Classical Autoencoders for End-to-End Radio Communication
CN113055029A (en) System polarization code encoding and decoding integrated device capable of multiplexing resources and encoding and decoding method
Tato et al. Deep learning assisted rate adaptation in spatial modulation links
CN113992313B (en) Balanced network assisted SCMA encoding and decoding method based on deep learning
CN109039531B (en) Method for adjusting LT code coding length based on machine learning
CN114759997B (en) MIMO system signal detection method based on data model double driving
CN113852434B (en) LSTM and ResNet-assisted deep learning end-to-end intelligent communication method and system
CN116405158A (en) End-to-end communication system based on deep learning under non-Gaussian noise
CN116249128A (en) Joint beam forming method based on end-to-end learning in multi-RIS auxiliary communication system
CN112821971A (en) Time-varying channel signal detection method based on countermeasure learning
Adali Why a nonlinear solution for a linear problem?[channel equalization]
CN114024803B (en) Channel estimation method under OFDM system

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