CN116346263A - Construction method and application of intelligent signal detection model - Google Patents

Construction method and application of intelligent signal detection model Download PDF

Info

Publication number
CN116346263A
CN116346263A CN202310123894.2A CN202310123894A CN116346263A CN 116346263 A CN116346263 A CN 116346263A CN 202310123894 A CN202310123894 A CN 202310123894A CN 116346263 A CN116346263 A CN 116346263A
Authority
CN
China
Prior art keywords
node
signal detection
intelligent signal
detection model
structure data
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
CN202310123894.2A
Other languages
Chinese (zh)
Inventor
张旭帆
江涛
肖丽霞
丁超
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.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and 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 Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202310123894.2A priority Critical patent/CN116346263A/en
Publication of CN116346263A publication Critical patent/CN116346263A/en
Pending legal-status Critical Current

Links

Images

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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/309—Measuring or estimating channel quality parameters
    • H04B17/336—Signal-to-interference ratio [SIR] or carrier-to-interference ratio [CIR]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0242—Channel estimation channel estimation algorithms using matrix methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • 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)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Power Engineering (AREA)
  • Quality & Reliability (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a construction method and application of an intelligent signal detection model, belonging to the technical field of mobile communication; the invention constructs a novel graph neural network comprising an aggregation updating module, and aggregates the global characteristics of each node and the neighbor node information thereof, thereby extracting the hidden characteristic state of the data according to the neighbor nodes and the characteristic information thereof, and further realizing the effective capture of the characteristics of non-Euclidean space data such as communication signals; in addition, after each polymerization is completed, the invention further performs feature extraction on the intermediate features obtained by the polymerization to improve the expression capability of the intermediate features, so that the designed graph neural network has more efficient information interaction capability and more powerful relationship reasoning capability, and the signal detection precision is greatly improved. The invention effectively solves the problem that the performance and the computational complexity of the traditional signal detection method are difficult to balance, and has strong robustness to channel estimation errors.

Description

Construction method and application of intelligent signal detection model
Technical Field
The invention belongs to the technical field of mobile communication, and particularly relates to a construction method and application of an intelligent signal detection model.
Background
The 6G mobile communication has the remarkable characteristics of generalization, socialization, intellectualization and the like, and the channel propagation range covers the complex environments such as ultra-large bandwidth, high, medium and low multi-frequency bands, space, land, sea and the like. The existing orthogonal frequency division multiplexing technology is difficult to realize high-performance transmission in a 6G communication complex channel environment, and development of a novel waveform modulation technology is needed.
However, existing signal detection methods face significant challenges in the face of new waveform modulation techniques. Taking orthogonal time-frequency space (Orthogonal Time Frequency Space, OTFS) as an example, the method is a novel waveform technology for obtaining multidimensional diversity by using delay-doppler domain, has lower peak-to-average ratio, higher spectral efficiency and stronger channel time-varying resistance, and is one of modulation technologies with great potential for 6G communication. Due to the doppler effect, the larger dimension of the equivalent channel matrix, and the like, the conventional signal detection method is difficult to directly apply. Specifically, the conventional linear detection method mainly performs interference cancellation by constructing an equalization matrix, and has low complexity but serious performance loss. However, the traditional nonlinear detection method, such as message transmission and expected propagation, has a detection performance which is closely related to the computational complexity, and it is difficult to realize high-precision detection of signals under the condition of low complexity.
On the other hand, with successful application of deep learning in fields of natural image processing, computer vision and the like, intelligent signal detection algorithms based on the technology are getting more and more attention. However, conventional deep learning techniques, such as deep neural networks (Deep Neural Networks, DNN), convolutional neural networks, etc., mainly address euclidean space data, and need to satisfy translational invariance. However, for the problem of communication signal detection, the variable nodes and the factor nodes naturally form a factor graph data model, which belongs to a typical two-part graph, and the translation invariance is not satisfied, so that the performance of the existing intelligent signal detection algorithm is limited, and the accuracy of a detection result is lower.
Disclosure of Invention
Aiming at the defects or improvement demands of the prior art, the invention provides a construction method and application of an intelligent signal detection model, which are used for solving the technical problem that the detection result precision of the existing communication signal detection method is lower.
In order to achieve the above object, in a first aspect, the present invention provides a method for constructing an intelligent signal detection model, including:
s1, acquiring a plurality of groups of communication system data comprising a transmission signal, a channel matrix and a receiving signal, and respectively converting the channel matrix and the receiving signal in each communication system data into corresponding graph structure data;
s2, training a graph neural network model by taking graph structure data as input and corresponding sending signals as output to obtain an intelligent signal detection model;
the graph neural network model comprises an aggregation updating module, a graph neural network model generation module and a graph neural network model generation module, wherein the aggregation updating module is used for carrying out depth aggregation on global features of each node and neighbor node information of each node through multiple iterations; in each iteration process, the Laplace operator is adopted to multiply the global features of each node in the previous iteration respectively, the intermediate features of each node are obtained, and then the intermediate features are input into a feature extraction unit for deep feature extraction, so that the global features of each node in the current iteration are obtained;
the initial global feature of each node is a result obtained by respectively carrying out convolution operation on each node parameter in the graph structure data; the node parameters are constructed based on the channel matrix, the received signal and the system noise figure.
Further preferably, the above-mentioned feature extraction unit includes: cascaded DNN modules and GRU units;
the DNN module is used for respectively inputting the node intermediate characteristics into K DNN networks, summing the results output by the DNN networks and outputting the summed results to the GRU unit; k is more than or equal to 2.
Further preferably, the DNN network is a three-layer DNN network; each DNN network contains two hidden layers.
Further preferably, the graph neural network model further includes: a convolution module and an output module;
the convolution module is used for carrying out convolution operation on each node parameter in the graph structure data respectively to obtain initial global characteristics of each node, and outputting the initial global characteristics to the aggregation updating module;
the output module is used for respectively carrying out feature mapping on the global features of each node output by the aggregation module to obtain the predicted value of the transmitting symbol corresponding to each node, and further obtaining the detection result of the transmitting signal.
Further preferably, the output module is a three-layer DNN network.
Further preferably, the laplace operator
Figure BDA0004081051460000031
Figure BDA0004081051460000032
Is a diagonal matrix>
Figure BDA0004081051460000033
Figure BDA0004081051460000034
Edge characteristic matrix of graph structure data, element e thereof ij ＝h i T ×h j ；h i Is the ith column of the channel matrix; c is->
Figure BDA0004081051460000035
Is a column number of columns.
Further preferably, the method for constructing graph structure data in step S1 includes: based on the channel matrix of each group of communication system data, the corresponding relation between the transmitting symbol and the receiving symbol is obtained, and the transmitting symbol is taken as a node, and the corresponding graph structure data is constructed by connecting all transmitting symbol nodes corresponding to the same receiving symbol in pairs.
Further preferably, for the ith node of the graph structure data, the node parameters thereof are:
f i ＝[y T h i ,h i T h i ,σ 2 ]
wherein y is a received signal; h is a i An ith column of channel vectors in the channel matrix;σ 2 is the system noise figure.
Further preferably, the step S1 includes: according to the configuration information of the communication system, a plurality of groups of sending signals x and channel matrixes H are randomly generated, and corresponding receiving signals y are obtained according to a communication system model y=Hx+n; where n is the system noise.
Through the simulation communication system, massive training data sets are directly and randomly generated, additional labels are not needed, sufficient training samples can be obtained more efficiently and at low cost, and the effectiveness of network training is ensured.
In a second aspect, the present invention provides an intelligent signal detection method, including: after a receiving end of the communication system acquires a receiving signal and a channel matrix, converting the receiving signal and the channel matrix into corresponding graph structure data; inputting the graph structure data into an intelligent signal detection model to obtain a detection result of a transmitted signal;
the intelligent signal detection model is constructed by adopting the construction method of the intelligent signal detection model provided by the first aspect of the invention.
In a third aspect, the present invention provides an intelligent signal detection system comprising: the intelligent signal detection system comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the intelligent signal detection method provided by the second aspect of the invention when executing the computer program.
In a fourth aspect, the present invention further provides a computer readable storage medium, where the computer readable storage medium includes a stored computer program, where the computer program, when executed by a processor, controls a device where the storage medium is located to execute the method for constructing the intelligent signal detection model provided in the first aspect of the present invention and/or the method for detecting an intelligent signal provided in the second aspect of the present invention.
In general, through the above technical solutions conceived by the present invention, the following beneficial effects can be obtained:
1. the invention provides a construction method of an intelligent signal detection model, which constructs a novel graph neural network comprising an aggregation update module, and aggregates global features of each node and neighbor node information thereof, so as to extract hidden feature states of data according to the neighbor nodes and the feature information thereof, and further realize effective capture of features of non-Euclidean space data such as communication signals; in addition, after each polymerization is completed, the invention further performs feature extraction on the intermediate features obtained by the polymerization to improve the expression capability of the intermediate features, so that the designed graph neural network has more efficient information interaction capability and more powerful relationship reasoning capability, and the signal detection precision is greatly improved.
2. According to the method for constructing the intelligent signal detection model, a plurality of DNN networks are adopted to process intermediate features aggregated with neighbor node information at the same time, and then the features obtained by the DNN networks are fused to obtain more comprehensive and more obvious node features, so that the expression capacity of the graph neural network is further improved; finally, inputting the fused characteristics into a GRU unit for processing; because some structures and context relations are hidden among all nodes (detection symbols) in signal detection, the context information contained among data and the hidden relevance among the context information can be further extracted by adopting the GRU (global packet transfer unit) circulating neural network module, and the propagation capability of long-term information (namely the history information and the long-term state stored by a memory unit) of a graph structure is further improved. By the method, the promotion effect of the graph neural network on signal detection is greatly improved, hidden features of nodes can be efficiently extracted at the cost of lower computational complexity and memory, and the accuracy of signal detection is further improved.
3. According to the method for constructing the intelligent signal detection model, communication system data are expressed into the graph structural form based on the incidence relation between the emission symbols and then are input into the graph neural network designed by the method for training, so that feature extraction can be better learned and performed by utilizing the relation between the emission symbols, node features are updated with lower calculation complexity and memory requirements, and finally a detection result with better performance is obtained.
4. The invention utilizes the strong relation reasoning and information interaction capability of the constructed graph neural network, effectively solves the problem that the performance and the computational complexity of the traditional signal detection method are difficult to balance, and has strong robustness on channel estimation errors; experiments show that the detection result obtained by the intelligent signal detection method based on the graph neural network is superior to that obtained by the traditional signal detection method, and has strong robustness on imperfect channel state information.
Drawings
FIG. 1 is a schematic diagram of the neural network model according to the present invention;
FIG. 2 is a schematic diagram of an OTFS system according to an embodiment of the present invention;
FIG. 3 is a flowchart of an intelligent signal detection method applied to an OTFS system according to an embodiment of the present invention;
fig. 4 is a code rate simulation diagram obtained by adopting different signal detection methods according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.
In the present invention, the terms "first," "second," and the like in the description and in the drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order.
In a first aspect, the present invention provides a method for constructing an intelligent signal detection model, including:
s1, acquiring a plurality of groups of communication system data comprising a transmission signal, a channel matrix and a receiving signal, and respectively converting the channel matrix and the receiving signal in each communication system data into corresponding graph structure data;
it should be noted that, when acquiring the data of the communication system, the communication system may be directly acquired in the actual sceneThe data in the communication system can also be directly and randomly generated into mass communication system data through the simulation communication system, and no additional tag is needed. In order to obtain sufficient training samples more efficiently and at low cost, and ensure the effectiveness of network training, data collection is preferably realized in a simulation manner, specifically, according to the configuration information of a communication system (including the number N of antennas at a transmitting end t Number N of receiving end antennas r The number of subcarriers M, the number of time slots N, the number of multipaths L, the center frequency f c Bandwidth B, maximum delay τ max Signal modulation mode, signal-to-noise ratio of channel, etc.), randomly generating multiple groups of transmission signals x and channel matrix H, and acquiring corresponding receiving signals y according to a communication system model y=hx+n; where n is the system noise, based on the signal noise figure sigma 2 And (5) calculating to obtain the product.
Further, when constructing the graph structure data, the invention obtains the corresponding relation between the transmitting symbol and the receiving symbol based on the channel matrix of each group of communication system data, and uses the transmitting symbol as a node, and constructs the corresponding graph structure data by connecting all transmitting symbol nodes corresponding to the same receiving symbol. Wherein the edge feature in the graph structure data is obtained based on the channel vectors corresponding to the two connected nodes, specifically, the edge feature e between the ith transmitting symbol and the node corresponding to the jth transmitting symbol ij ＝h i T ×h j ；h i Is the ith column of the channel matrix. The node parameters in the graph structure data are constructed based on the channel matrix, the received signal and the system noise coefficient. In an alternative, the conventional signal detection results may be used as initial characteristics of the nodes and then optimized using the graph network. For example, the detection result of the approach zero algorithm or the minimum mean square error algorithm may be used as the initial feature of the node. For example, when the approach zero algorithm is adopted, node parameters of the ith node of the graph structure data are: f (f) i ＝[H -1 y i ]The method comprises the steps of carrying out a first treatment on the surface of the When the minimum mean square error algorithm is adopted, node parameters of the ith node of the graph structure data are as follows: f (f) i ＝H H (H H +σ 2 I) -1 y i The method comprises the steps of carrying out a first treatment on the surface of the Wherein y is i For the ith received symbol in the received signal; sigma (sigma) 2 Is the system noise figure.
Further, it is considered that the conventional linear signal detection algorithm is to multiply the received signal y by constructing various equalization matrices to obtain a signal detection value x. And y is T h，h T h and noise figure are key parameters for constructing these equalization matrices, so in order to highlight these key parameters, a better detection result is obtained, and in another alternative, for the ith node of the graph structure data, the node parameters are:
f i ＝[y T h i ,h i T h i ,σ 2 ]
wherein y is a received signal; h is a i An ith column of channel vectors in the channel matrix; sigma (sigma) 2 Is the system noise figure.
It should be noted that, when the communication system (the expression of the model y=hx+n) is given, the transmission symbol x can be obtained based on the channel matrix H i And receiving symbol y j The factor graph (corresponding relation between the transmitting symbol and the receiving symbol) between the two symbols can be known and deduced by taking the receiving symbol as an intermediate node; the invention fully utilizes the structural relation of the communication system, converts the factor graph containing two types of nodes of the transmitting symbol and the receiving symbol into the graph structure containing only the nodes of the transmitting symbol, and is more beneficial to the designed graph neural network to analyze the association information between the receiving symbols. According to the invention, based on the association relation between the transmitting symbols, the communication system data is expressed into a graph structural form and then is input into the designed graph neural network for training, so that the characteristic extraction can be better learned and carried out by utilizing the relation between the transmitting symbols, the updating of the node characteristics is realized with lower calculation complexity and memory requirement, and finally, the detection result with better performance is obtained.
S2, training a graph neural network model by taking graph structure data as input and corresponding sending signals as output to obtain an intelligent signal detection model;
the neural network model comprises a cascade convolution module, an aggregation updating module and an output module as shown in fig. 1;
1) And a convolution module:
the convolution module is used for carrying out convolution operation on each node parameter in the graph structure data respectively to obtain initial global characteristics of each node, and outputting the initial global characteristics to the aggregation updating module; in an alternative implementation mode, each node parameter in the graph structure data is encoded through a single-layer full-connection layer respectively, and initial global characteristics of each node are obtained; taking the construction mode of the second node parameter as an example, the initial global feature of the ith node is:
Figure BDA0004081051460000081
wherein W is 1 Is of dimension R n Learner parameter matrix of x 3 single-layer network, b 1 Is a bias vector that it can learn.
2) An aggregation updating module:
the aggregation updating module is used for carrying out depth aggregation on the global characteristics of each node and the neighbor node information of each node through multiple iterations, and comprises an aggregation unit and a characteristic extraction unit; in each iteration process, the aggregation unit multiplies the global features of each node under the previous iteration by using the Laplacian to obtain intermediate features of each node, and then inputs the intermediate features into the feature extraction unit for deep feature extraction, so that the global features of each node under the current iteration are obtained;
specifically, the intermediate features of the ith node at the nth iteration
Figure BDA0004081051460000082
Wherein u is i (t-1) Global features of the ith node under the t-1 th iteration; laplacian>
Figure BDA0004081051460000083
Figure BDA0004081051460000087
In the form of a diagonal matrix,
Figure BDA0004081051460000084
Figure BDA0004081051460000085
edge characteristic matrix of graph structure data, element e thereof ij ＝h i T ×h j ；h i Is the ith column of the channel matrix; c is->
Figure BDA0004081051460000086
Is a column number of columns.
In an alternative embodiment, the total number of iterations in the aggregate update module is 10.
In order to further promote the promotion effect of the neural network on the signal detection algorithm, a concept of 'multi-head' is introduced in an alternative implementation mode, firstly, a plurality of DNN networks are utilized to process intermediate features obtained after aggregation with neighbor node information, then more remarkable node features are obtained through feature fusion, and finally, the expression capacity of the graph neural network can be further improved through GRU units. Specifically, the above-described feature extraction unit includes: cascaded DNN modules and GRU units; the DNN module is used for respectively inputting the node intermediate characteristics into K DNN networks, summing the results output by the DNN networks and outputting the summed results to the GRU unit; k is more than or equal to 2. In a preferred embodiment, the DNN network is a three-layer DNN network; each DNN network comprises two hidden layers, each comprising N 1 And N 2 The number of the neurons of the output layer is R n . Furthermore, the activation functions of both the hidden layer and the output layer are ReLU functions.
The input of GRU unit is node characteristic d after current fusion processing (i.e. after summation of K DNN network output results) i (t) And the node characteristic d after the fusion processing in the last iteration is reserved i (t-1) The output is dimension R n New feature of x 1.
Further, the above feature extraction unit further includes: and the scale change unit is used for transforming the feature sizes output by the GRU unit so as to keep the sizes of the global features obtained under each iteration consistent.
3) And an output module:
the output module is used for respectively carrying out feature mapping on the global features of each node output by the aggregation module to obtain the predicted value of the transmitting symbol corresponding to each node, and further obtaining the detection result of the transmitting signal. In a preferred embodiment, the output module is a three-layer DNN network, which includes two hidden layers, and uses a ReLU function as its corresponding activation function; the number of neurons of the output layer is determined by the modulation mode of the communication system, and the activation function is a softmax function.
The invention uses a small DNN network as an output module, and can effectively classify the nodes at the cost of smaller calculation complexity to realize signal detection. Test results show that the output module can obtain good detection precision.
In summary, the invention constructs a novel graph neural network with high-efficiency information interaction and strong relation reasoning capability as an intelligent signal detector, and deep-aggregates the global feature of each node and the neighbor node information thereof, so that the hidden feature state of the data can be extracted according to the neighbor node and the feature information thereof, thereby realizing effective capturing of the features of non-Euclidean space data such as communication signals, having higher signal detection accuracy, balancing the performance and the computational complexity of signal detection, and having strong robustness to channel estimation errors by virtue of the deep-learning high-efficiency nonlinear representation capability.
In a second aspect, the present invention provides an intelligent signal detection method, including:
after a receiving end of the communication system acquires a receiving signal and a channel matrix, converting the receiving signal and the channel matrix into corresponding graph structure data; inputting the graph structure data into an intelligent signal detection model to obtain a detection result of a transmitted signal;
the intelligent signal detection model is constructed by adopting the construction method of the intelligent signal detection model provided by the first aspect of the invention.
It should be noted that, when performing intelligent signal detection, the method for converting the received signal and the channel matrix into the corresponding graph structure data is the same as the conversion method adopted in the method for constructing the intelligent signal detection model provided in the first aspect of the present invention. The related technical solution is the same as the method for constructing the intelligent signal detection model provided in the first aspect of the present invention, and will not be described herein.
In a third aspect, the present invention provides an intelligent signal detection system comprising: the intelligent signal detection system comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the intelligent signal detection method provided by the second aspect of the invention when executing the computer program.
The related technical solution is the same as the intelligent signal detection method provided in the second aspect of the present invention, and details are not described here.
In a fourth aspect, the present invention further provides a computer readable storage medium, where the computer readable storage medium includes a stored computer program, where the computer program, when executed by a processor, controls a device where the storage medium is located to execute the method for constructing the intelligent signal detection model provided in the first aspect of the present invention and/or the method for detecting an intelligent signal provided in the second aspect of the present invention.
The related technical solution is the same as the method for constructing the intelligent signal detection model provided in the first aspect and the method for detecting an intelligent signal provided in the second aspect, which are not described herein.
In order to further illustrate the technical solution provided by the present invention, the following details are described in connection with specific embodiments:
the existing intelligent detection method mainly relies on the traditional deep learning technology and mainly aims at Euclidean space data. In the communication signal detection problem, the variable nodes and the factor nodes form a factor graph relation, and the translation invariance is not satisfied, so that the performance of the existing intelligent detection method is degraded. Therefore, the embodiment provides an intelligent signal detection method based on a graph neural network, which is used for an OTFS system, and can effectively balance signal detection performance and calculation complexity, thereby realizing efficient detection of OTFS signals.
Consider a oneSuch as the OTFS system model shown in fig. 2. The information bits are first amplitude phase modulated to produce M x N constellation symbols, where M is the number of subcarriers and N is the slot size. Then it is converted from the delay-doppler domain into time-frequency domain symbols by a two-dimensional inverse-octave fourier transform
Figure BDA0004081051460000111
Processing the time-frequency domain signal by utilizing the Hessenberg transformation and the window function to obtain a transmission signal +.>
Figure BDA0004081051460000112
After the above processing, the signal s (t) is wirelessly transmitted in a time-varying channel, and the receiving end receives the signal r (t) = ≡h (τ, v) s (t- τ) e j2πv(t-τ) dτdv. Where h (τ, v) represents the impulse response of the channel and τ and v represent the delay and doppler frequency offsets, respectively. Then, carrying out the wiener transformation on r (t) to obtain a time-frequency domain signal
Figure BDA0004081051460000113
Wherein (1)>
Figure BDA0004081051460000114
Is a matching window function of the receiving end. Finally, the time-frequency domain signal Y is converted by utilizing the Fourier transform mn The transition is to the delay-doppler domain,
Figure BDA0004081051460000115
the corresponding matrix form of the receiver portion of the system can then be expressed as:
Figure BDA0004081051460000116
wherein, the liquid crystal display device comprises a liquid crystal display device, F n and->
Figure BDA0004081051460000117
Fast fourier transform and inverse fourier transform, respectively representing N points,>
Figure BDA0004081051460000118
is Cronecker product, H tv Is a matrix of radio channels, G rx And G tx Diagonal matrices are constructed for the receive and transmit end window functions, respectively. Here->
Figure BDA0004081051460000121
Equivalent channel H denoted OTFS eff ，/>
Figure BDA0004081051460000122
Is equivalent noise->
Figure BDA0004081051460000123
Detection of OTFS signals can be equivalently solved as follows:
Figure BDA0004081051460000124
where p (y|x) =p y|z (y|H eff X), and z=h eff ·x。
The embodiment of the invention applies an intelligent signal detection method based on a graph neural network to the OTFS system, and the whole flow is shown in a figure 3, and specifically comprises the following steps:
step one: offline training phase
S1, generating a receiving signal matrix and an equivalent channel matrix as input of a training data set according to OTFS system configuration information, and generating a transmitting signal matrix as a label of the training data set.
Specifically, firstly according to the configuration information of the OTFS system, the method includes: number of subcarriers M, number of slots N, number of paths L, carrier frequency f c Maximum delay index
Figure BDA0004081051460000125
Doppler index Range->
Figure BDA0004081051460000126
Relative velocity v between transmitting and receiving means c The signal modulation scheme and the signal-to-noise ratio of the communication system channel randomly generate the transmission signal. Then, according to the formula
Figure BDA0004081051460000127
And calculating an OTFS equivalent channel matrix. Then according to the established system model
Figure BDA0004081051460000128
Calculating a receiving end signal y from the input signal, the equivalent channel matrix and the system noise, wherein
Figure BDA0004081051460000129
Is the corresponding equivalent noise. Finally, the received signal y and the equivalent channel matrix H eff As input to the training data set and the transmitting end signal x as a tag.
S2, constructing a graph neural network model, which mainly comprises three functional modules, namely aggregation, updating and output; and the user of the OTFS system is used as a node of the graph neural network, and the initial characteristic information of the node is sequentially processed by the three modules to obtain a signal detection result.
Specifically, the graph neural network built by the invention combines the advantages of deep learning and a Markov random field, captures the structural information of data as a feature vector, and updates the feature vector through message transmission among nodes. A set of random variables x= { x 0 ，...，x N-1 The structural signal of } can be represented by an undirected graph
Figure BDA0004081051460000131
Modeling is performed in which V and E represent sets of nodes and edges, respectively. Each node corresponds to a random variable one by one and satisfies p (x i |x\x i )＝p(x i I ne (i)), where \ refers to the exclude operation, ne (i) is the neighbor node set of the i-th node. In a pairwise relationship of Markov random fields, the self-potential +.>
Figure BDA0004081051460000132
Assigned to the ith node and will pairVigor->
Figure BDA0004081051460000133
And assigning an edge of the ith node connected with the jth node. The corresponding posterior probability can be expressed as +.>
Figure BDA0004081051460000134
Wherein C is a normalized constant and +.>
Figure BDA0004081051460000135
Wherein sigma 2 Is the noise figure of the OTFS system, h i Is an equivalent channel matrix H eff I th column, p i (x i ) Is the a priori probability of the variable. The graph neural network is mainly used for pushing the posterior probability p through nodes and edges GNN (x|y), and self-potential and counter-potential are taken as feature vectors of nodes and edges.
In the present embodiment use
Figure BDA0004081051460000136
To define an initial eigenvector of the ith node, which is defined by the received signal y, the ith column channel matrix h i And system noise figure sigma 2 And (5) jointly determining. The node parameters are encoded through a single-layer full-connection layer to obtain the initial characteristic of the node ∈>
Figure BDA0004081051460000137
And serves as an initial global feature of the node; wherein W is 1 Is that the dimension of the single-layer network is R n X 3 matrix of learnable parameters, b 1 Is a learnable parameter vector. In addition, the characteristics of the edge between the ith node and the jth node may be expressed as +.>
Figure BDA0004081051460000138
And then, the node characteristics and the edge characteristics are sequentially processed by an aggregation updating module and an output module of the graph network, so that a final detection result is obtained.
Specifically, the polymerization process first calculates the correspondingLaplacian operator
Figure BDA0004081051460000139
Wherein->
Figure BDA00040810514600001310
Is a diagonal matrix, and->
Figure BDA00040810514600001311
Then, information between nodes is aggregated by a graph convolution operation based on the calculated Laplace operator>
Figure BDA00040810514600001312
In the formula->
Figure BDA00040810514600001313
Is a node characteristic matrix (intermediate characteristic) after aggregation of neighbor node information under the t-th iteration; in the first iteration, ∈>
Figure BDA0004081051460000141
For initial global feature->
Figure BDA0004081051460000142
In a subsequent iteration of the process,
Figure BDA0004081051460000143
the result is output by the feature extraction unit under the t-1 th iteration.
Specifically, the feature extraction unit is composed of K three-layer DNN networks and a GRU unit (gating recursion unit) together. Wherein each DNN network comprises two hidden layers, respectively comprising N 1 And N 2 The number of the output layer nerves of each neuron is R n . Furthermore, the activation functions of both the hidden layer and the output layer are ReLU. In order to further promote the promotion effect of the neural network on the signal detection algorithm, the invention introduces the concept of multiple heads in the graph-meaning network, namely, the K DNN networks are utilized to process the aggregated node characteristics at the same time, and then the node characteristics are fused to obtain a more comprehensive new nodeFeatures. And then the GRU unit is utilized to improve the expression capacity of the graph neural network. The input of the module is the node characteristics processed by K DNN networks at present
Figure BDA0004081051460000144
And the characteristics of the node in the last iteration +.>
Figure BDA0004081051460000145
The output is dimension R n New feature of x 1.
Specifically, the output module is composed of a three-layer DNN module. The system comprises two hidden layers, and uses a ReLU function as a corresponding activation function, the number of neurons of an output layer is determined by an OTFS system modulation mode, and the activation function is a softmax function.
S3, offline training is carried out on the built network by utilizing the randomly generated training set data, and the weight parameters of the network are adjusted.
Specifically, the offline training process comprises the core steps of generating a training data set, configuring network super parameters, setting a loss function and the like. Wherein the training data set is randomly generated (x, H) eff ，σ 2 Y) composition. Receiving signal y, effective channel matrix H eff And system noise figure sigma 2 Used as input to the training dataset and x as the label. Each training cycle contains 30000 sets of training data, with 24000 sets of data being used as training sets and the remaining 6000 sets of data being used as validation sets, with the batch size set to 100. In addition, the super parameters of the network are configured as follows: the number of loops of the graph network is set to 10, and the number of neurons in the network is R n ＝8，N 1 ＝32，N 2 =16, multi-head number k=3, learning rate 1×10 -4 The total number of cycles trained was 3000. Finally, the invention selects the cross entropy as the corresponding loss function:
Figure BDA0004081051460000151
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure BDA0004081051460000152
is the final output of the graphic neural network, p s (x i ) Is the x true probability distribution.
And calculating a corresponding gradient through the obtained loss function value in each training, updating the weight and the bias parameters of the neural network of the graph according to the selected gradient descent rule, and storing the optimal weight value and the optimal bias parameters.
The embodiment records the weight parameter and the bias parameter of each training of the network and stores the relevant parameters corresponding to the optimal training result. Through the set network super parameters such as batch size, learning rate, training cycle times, loss function and the like, the high-efficiency training of the network is realized, and the convergence and the effectiveness of the network are ensured.
Step two: on-line detection stage
S4, carrying out online detection on the information obtained by the OTFS system receiving end by utilizing the trained network to obtain an actual detection result of the network.
The following further verifies the beneficial effects of the signal detection method based on the graph neural network provided by the embodiment of the invention compared with the prior art by using the existing minimum mean square error based on the message transmission signal detection method as a comparison. In fig. 4, the abscissa represents the signal-to-noise Ratio (snr), and the ordinate represents the Bit Error Rate (BER), with smaller values representing better performance of the corresponding algorithm. Fig. 3 provides the relationship between the error rate and the signal-to-noise ratio obtained by the signal detection method in fig. 3, where: the invention provides a signal detection method based on a graph neural network, which comprises the steps of minimum mean square error signal detection and message transmission signal detection.
In fig. 4, the parameter setting condition is that the OTFS system works in the 4GHz band, the number of subcarriers m=8, the number of slots n=8, the number of paths l=4, the relative speed between the transmitting device and the receiving device is 150Kmph, and the system adopts a QPSK modulation mode. The signal detection method based on the graph neural network is far superior to the other two traditional methods, and the advantages are gradually enlarged along with the increase of the signal-to-noise ratio.
It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims (10)

1. The method for constructing the intelligent signal detection model is characterized by comprising the following steps of:
s1, acquiring a plurality of groups of communication system data comprising a transmission signal, a channel matrix and a receiving signal, and respectively converting the channel matrix and the receiving signal in each communication system data into corresponding graph structure data;
s2, training a graph neural network model by taking graph structure data as input and corresponding sending signals as output to obtain an intelligent signal detection model;
the graph neural network model comprises an aggregation updating module, a processing module and a processing module, wherein the aggregation updating module is used for carrying out depth aggregation on the global characteristics of each node and the neighbor node information of each node through multiple iterations; in each iteration process, the Laplace operator is adopted to multiply the global features of each node in the previous iteration respectively, the intermediate features of each node are obtained, and then the intermediate features are input into a feature extraction unit for deep feature extraction, so that the global features of each node in the current iteration are obtained;
the initial global feature of each node is a result obtained by respectively carrying out convolution operation on each node parameter in the graph structure data; the node parameters are constructed based on a channel matrix, a received signal and a system noise coefficient.
2. The method for constructing an intelligent signal detection model according to claim 1, wherein the feature extraction unit comprises: cascaded DNN modules and GRU units;
the DNN module is used for respectively inputting node intermediate features into K DNN networks, summing the results output by the DNN networks and outputting the summed results to the GRU unit; k is more than or equal to 2.
3. The method for constructing an intelligent signal detection model according to claim 1, wherein the graph neural network model further comprises: a convolution module and an output module;
the convolution module is used for carrying out convolution operation on each node parameter in the graph structure data respectively to obtain initial global characteristics of each node, and outputting the initial global characteristics to the aggregation updating module;
the output module is used for respectively carrying out feature mapping on the global features of the nodes output by the aggregation module to obtain predicted values of the transmitting symbols corresponding to the nodes, and further obtaining detection results of the transmitting signals.
4. The method for constructing an intelligent signal detection model according to claim 1, wherein the laplace operator
Figure FDA0004081051440000021
Figure FDA0004081051440000022
Is a diagonal matrix>
Figure FDA0004081051440000023
Figure FDA0004081051440000024
Edge characteristic matrix of graph structure data, element e thereof ij ＝h i T ×h j ；h i Is the ith column of the channel matrix; c is->
Figure FDA0004081051440000025
Is a column number of columns.
5. The method for constructing an intelligent signal detection model according to any one of claims 1 to 4, wherein the method for constructing graph structure data in step S1 includes: based on the channel matrix of each group of communication system data, the corresponding relation between the transmitting symbol and the receiving symbol is obtained, and the transmitting symbol is taken as a node, and the corresponding graph structure data is constructed by connecting all transmitting symbol nodes corresponding to the same receiving symbol in pairs.
6. The method for constructing an intelligent signal detection model according to any one of claims 1 to 4, wherein for an i-th node of the graph structure data, node parameters thereof are:
f i ＝[y T h i ,h i T h i ,σ 2 ]
wherein y is a received signal; h is a i An ith column of channel vectors in the channel matrix; sigma (sigma) 2 Is the system noise figure.
7. The method for constructing an intelligent signal detection model according to any one of claims 1 to 4, wherein the step S1 includes: according to the configuration information of the communication system, a plurality of groups of sending signals x and channel matrixes H are randomly generated, and corresponding receiving signals y are obtained according to a communication system model y=Hx+n; where n is the system noise.
8. An intelligent signal detection method is characterized by comprising the following steps: after a receiving end of the communication system acquires a receiving signal and a channel matrix, converting the receiving signal and the channel matrix into corresponding graph structure data; inputting the graph structure data into an intelligent signal detection model to obtain a detection result of a transmitted signal;
the intelligent signal detection model is constructed by adopting the construction method of the intelligent signal detection model according to any one of claims 1-7.
9. An intelligent signal detection system, comprising: a memory storing a computer program and a processor that when executing the computer program performs the intelligent signal detection method of claim 8.
10. A computer readable storage medium, characterized in that the computer readable storage medium comprises a stored computer program, wherein the computer program, when being executed by a processor, controls a device in which the storage medium is located to perform the method of constructing the intelligent signal detection model according to any one of claims 1-7 and/or the intelligent signal detection method according to claim 8.
CN202310123894.2A 2023-02-16 2023-02-16 Construction method and application of intelligent signal detection model Pending CN116346263A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310123894.2A CN116346263A (en) 2023-02-16 2023-02-16 Construction method and application of intelligent signal detection model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310123894.2A CN116346263A (en) 2023-02-16 2023-02-16 Construction method and application of intelligent signal detection model

Publications (1)

Publication Number Publication Date
CN116346263A true CN116346263A (en) 2023-06-27

Family

ID=86876537

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310123894.2A Pending CN116346263A (en) 2023-02-16 2023-02-16 Construction method and application of intelligent signal detection model

Country Status (1)

Country Link
CN (1) CN116346263A (en)

Similar Documents

Publication Publication Date Title
Pham et al. Intelligent radio signal processing: A survey
CN113472706B (en) MIMO-OFDM system channel estimation method based on deep neural network
CN110336594B (en) Deep learning signal detection method based on conjugate gradient descent method
Yang et al. Deep learning aided method for automatic modulation recognition
Abdel‐Moneim et al. A survey of traditional and advanced automatic modulation classification techniques, challenges, and some novel trends
CN109474388B (en) Low-complexity MIMO-NOMA system signal detection method based on improved gradient projection method
CN112910811B (en) Blind modulation identification method and device under unknown noise level condition based on joint learning
Zhang et al. Modulated autocorrelation convolution networks for automatic modulation classification based on small sample set
Daldal et al. Deep long short-term memory networks-based automatic recognition of six different digital modulation types under varying noise conditions
CN103888391A (en) Signal blind detection method based on double Sigmoid chaotic neural network
Chen et al. EMD and VMD empowered deep learning for radio modulation recognition
CN114745233B (en) Joint channel estimation method and device based on pilot frequency design
Enku et al. Deep-learning based signal detection for MIMO-OTFS systems
Amin et al. Orthogonal least squares based complex-valued functional link network
Le Magoarou Similarity-based prediction for channel mapping and user positioning
Zhang et al. Graph neural network assisted efficient signal detection for OTFS systems
CN111313943A (en) Three-dimensional positioning method and device under deep learning assisted large-scale antenna array
Zhang et al. Efficient residual shrinkage CNN denoiser design for intelligent signal processing: Modulation recognition, detection, and decoding
Rahman et al. HyDNN: A hybrid deep learning framework based multiuser uplink channel estimation and signal detection for NOMA-OFDM system
CN112564830B (en) Deep learning-based dual-mode orthogonal frequency division multiplexing index modulation detection method and device
Zhang et al. Toward the Automatic Modulation Classification With Adaptive Wavelet Network
CN116346263A (en) Construction method and application of intelligent signal detection model
CN114422310B (en) Digital quadrature modulation signal identification method based on joint distribution matrix and multi-input neural network
CN114759997B (en) MIMO system signal detection method based on data model double driving
CN116383656A (en) Semi-supervised characterization contrast learning method for large-scale MIMO positioning

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