CN114528925B - Time-varying channel OFDM signal equalization method based on deep classification network - Google Patents
Time-varying channel OFDM signal equalization method based on deep classification network Download PDFInfo
- Publication number
- CN114528925B CN114528925B CN202210109366.7A CN202210109366A CN114528925B CN 114528925 B CN114528925 B CN 114528925B CN 202210109366 A CN202210109366 A CN 202210109366A CN 114528925 B CN114528925 B CN 114528925B
- Authority
- CN
- China
- Prior art keywords
- classification
- frequency domain
- time
- signal
- layer
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000013135 deep learning Methods 0.000 claims abstract description 49
- 238000010586 diagram Methods 0.000 claims abstract description 47
- 238000012545 processing Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims abstract description 16
- 238000004891 communication Methods 0.000 claims abstract description 15
- 238000005259 measurement Methods 0.000 claims abstract description 6
- 230000000903 blocking effect Effects 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 50
- 230000005540 biological transmission Effects 0.000 claims description 33
- 239000011159 matrix material Substances 0.000 claims description 33
- 230000004913 activation Effects 0.000 claims description 25
- 230000014509 gene expression Effects 0.000 claims description 18
- 230000004044 response Effects 0.000 claims description 11
- 239000000654 additive Substances 0.000 claims description 10
- 230000000996 additive effect Effects 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 10
- 230000008054 signal transmission Effects 0.000 claims description 5
- 230000010363 phase shift Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 238000013459 approach Methods 0.000 claims description 2
- 238000003874 inverse correlation nuclear magnetic resonance spectroscopy Methods 0.000 claims description 2
- 230000009466 transformation Effects 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 description 11
- 230000000694 effects Effects 0.000 description 4
- 238000004088 simulation Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- 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
-
- 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
- G06N3/047—Probabilistic or stochastic 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
-
- 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
- H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
- H04L25/03006—Arrangements for removing intersymbol interference
- H04L25/03159—Arrangements for removing intersymbol interference operating in the frequency domain
-
- 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
- H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
- H04L25/03878—Line equalisers; line build-out devices
-
- 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)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Power Engineering (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention discloses a time-varying channel OFDM signal equalization method based on a deep classification network. The method comprises the following steps: given an actual measurement channel data set, constructing a deep learning network based on a classification idea, and respectively constructing input and output of the deep learning network based on the classification idea; performing online training on the constructed deep learning network based on the classification thought to obtain a deep classification network model; in specific time-varying channel communication, input data to be balanced are subjected to blocking processing, a deep learning network based on a classification idea is input, an available deep classification network model is called, and a single-heat coding estimated value of a frequency domain transmitting signal in a constellation diagram is obtained, so that time-varying channel OFDM signal balance under different modulation modes is realized. The invention utilizes the communication theory model and the diversity of the communication data set, and the network has better robustness to different actually measured channels, can adapt to channels with different complexity degrees and achieves better performance.
Description
Technical Field
The invention relates to an equalization technology in the field of communication, in particular to a time-varying channel OFDM signal equalization method based on a deep classification network.
Background
Equalization techniques eliminate the distortion effects by means of structural knowledge of the channel. The equalizer generally knows the channel impulse response or the impulse response of an incompletely known channel. The equalizer functions primarily to correct channel distortions to help recover the frequency domain transmitted signal. The output of the equalizer is typically a symbol sequence estimate. Since channel objects studied in the field of mobile communication are gradually expanding to the field of complex channels, the study of equalization techniques has great significance.
The conventional equalization method commonly used at present has minimum mean square error equalization (Minimum mean square error, MMSE) equalization and decision feedback (Decision feedback equalizer, DFE) equalization. The advantages of the deep learning (DEEP LEARNING, DL) algorithm in the field of communications have also gradually emerged in recent years. The current DL methods are largely divided into two main categories. One is to use a purely data driven network. In 2018, ye proposes an end-to-end OFDM system signal estimation model based on a DNN network. In this model, the channel estimation, signal equalization and demodulation modules of the OFDM receiver are all replaced by DNN networks. However, it does not use the mathematical relationship inherent between the transmitted and received signals in wireless communication, and loses a priori knowledge, so that the receiver is a black box, and the operation mechanism of the network is difficult to explain. In addition, the pure data driving method needs to rely on a large amount of data for training, and the general parameters are huge, so that the convergence speed is low and the calculation cost is high. The other is a model-and data-driven based network. The method based on the model and the data can combine the prior knowledge, and based on the classical communication theory, the deep learning network can replace the function of a certain module of the OFDM receiver. In 2019, the SCN equalizer combines expert knowledge of a wireless communication algorithm, makes estimation of a constellation diagram based on a regression idea, and is a deep regression network with poor robustness to different channels.
In the fast time-varying channel, the method has the characteristics of serious Doppler effect, time-varying multipath and the like. Orthogonal frequency division multiplexing (Orthogonal frequency division multiplexing, OFDM) techniques are widely used in such channels. The doppler shift generated by the fast time-varying channel may cause serious Inter-carrier interference (Inter-CARRIER INTERFERENCE, ICI) to occur in the OFDM signal, and reduce the performance of the conventional equalizer, thereby affecting the transmission of the OFDM signal. It is therefore important to develop OFDM signal equalization techniques that can cope with time-varying channels to reduce symbol error rates (Symbol error rate, SER).
Disclosure of Invention
The invention aims to overcome the defects and shortcomings of the prior art and provides a time-varying channel OFDM signal equalization method based on a classification deep learning network. Since constellation estimation is basically a multi-classification task, classification-based ideas perform better than deep learning networks built based on regression ideas. The equalization method based on the classified deep learning network can overcome the influence of time-varying channel distortion, obtain a more accurate single-hot (one-hot) coding estimated value of a frequency domain transmitting signal in a constellation diagram, and obtain better SER performance.
The object of the invention is achieved by at least one of the following technical solutions.
A time-varying channel OFDM signal equalization method based on deep classification network includes the following steps:
S1, given an actual measurement channel data set, constructing a deep learning network based on a classification idea, and respectively constructing input and output of the deep learning network based on the classification idea;
S2, performing online training on the constructed deep learning network based on the classification thought to obtain a deep classification network model;
S3, in specific time-varying channel communication, the input data to be balanced are subjected to blocking processing, a deep learning network based on a classification idea is input, an available deep classification network model is called, a single-heat coding estimated value of a frequency domain transmitting signal in a constellation diagram is obtained, and time-varying channel OFDM signal equalization in different modulation modes is realized.
Further, in step S1, the OFDM signal has N subcarriers, S (m) represents a signal transmitted on the mth subcarrier, and m=1 to N;
Where N represents the nth sampling time, n=0,..n-1; s (N) represents a time domain transmission signal after inverse fast fourier transform (INVERSE FAST fourier transform, IFFT) transformation at an nth sampling time N-point;
The time domain received signal y (n) of the OFDM time domain transmission signal s (n) arriving at the receiving end through the l paths at the nth sampling time is expressed as:
Where h (n, L) is the channel of the path of the nth sampling instant L, L is the total number of signal transmission paths, and z (n) is the time domain additive noise of the nth sampling instant;
the measured channel data set is a large number of actually measured channel sets; selecting X channels with equal length from the measured channel data set Obtaining a time-varying channel impulse response by using a linear interpolation mode, wherein the expression is as follows:
Where ζ is a coefficient of linear interpolation, ir is a time-varying channel impulse response generated using the measured channel dataset, x represents the number of channels selected, Represents the X-th selected channel, where x=1, …, X.
Further, performing two-dimensional fourier transform on the time-varying channel impulse response ir to obtain a time-varying frequency domain channel matrix H, which is specifically as follows:
the time-varying channel signal transmission process is modeled as:
Y=HS+Z;
wherein, Is a frequency domain received signal obtained by subjecting the time domain received signal y (n) to FFT processing,Is a frequency domain transmission signal obtained by FFT processing of a time domain transmission signal s (n)/>Is the frequency domain additive noise obtained after the time domain additive noise z (n) is subjected to FFT processing.
Further, quadrature Phase Shift Keying (QPSK) modulation or 16 Quadrature amplitude modulation (16-Quadrature Amplitude Modulation, 16-QAM) is performed on the frequency domain transmission signal S;
and performing single-heat coding on the frequency domain transmitting signals after QPSK or 16-QAM modulation, wherein the single-heat coding is respectively expressed as follows:
or=1,oi=1→[1,0,0,0]
or=-1,oi=1→[0,1,0,0]
or=1,oi=-1→[0,0,1,0]
or=-1,oi=-1→[0,0,0,1]
or=-3,oi=-3→[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=-1→[0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=3→[0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=1→[0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=-3→[0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=-1→[0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=3→[0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0]
or=-1,oi=1→[0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0]
or=3,oi=-3→[0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0]
or=3,oi=-1→[0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0]
or=3,oi=3→[0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0]
or=3,oi=1→[0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0]
or=1,oi=-3→[0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0]
or=1,oi=-1→[0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0]
or=1,oi=3→[0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0]
or=1,oi=1→[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1]
Where o r is the real part of the frequency domain transmitted signal after QPSK or 16-QAM modulation, and o i is the imaginary part of the frequency domain transmitted signal after QPSK or 16-QAM modulation; and performing single-heat coding on the frequency domain transmitting signal after QPSK or 16-QAM modulation to obtain a single-heat coded tag value P of the frequency domain transmitting signal in a constellation diagram.
Further, in step S1, the deep learning network based on the classification idea includes K classification network elements of the same structure stacked, where the kth classification network element sequentially includes an input layer, a first fully-connected layer, an activation function layer, a second fully-connected layer, and a classification layer, k=1, …, K.
Further, in the kth classifying network unit, the input layer receives the input frequency domain received signal Y, the time-varying frequency domain channel matrix H and the estimated value of the frequency domain transmitted signal in the kth-1 classifying network unit in the constellation diagramAnd outputting the spliced input matrix to the first full-connection layer, wherein the input matrix is expressed as:
Wherein λ k1 is a first step size parameter in the kth classified network element, λ k2 is a second step size parameter in the kth classified network element; when k=1, the initial estimated value of the frequency domain transmission signal in the constellation in the 1 st classified network element is input Initializing based on Zero Forcing (ZF) equalization method to obtain and optimize network performance; since the time-varying frequency domain channel matrix H is not necessarily a non-singular matrix, the Zero-forcing (ZF) -based equalization expression is in pseudo-inverse form
In a first fully connected layer, training and updating network weight parameters, the first fully connected layer is expressed as:
Wherein d k k-th classified network element is the output of the first full-connection layer, w k is the weight matrix of the k-th classified network element, and n k is the bias term of the k-th classified network element;
The activation function adopted in the activation function layer comprises relu functions or tanh functions, and the expressions are respectively:
relu(dk)=max(dk,0)
The output of the activation function layer is that the value A k=θk(dk),θk of the output d k of the first full-connection layer, which is subjected to nonlinear processing by the activation function, represents the activation function used by the kth classification network unit;
Output A k of the second full-connection layer input activation function layer outputs the estimated value of the frequency domain transmitting signal in the constellation diagram in the kth classification network unit
In the classification layer, according to the output A k of the input activation function layer, the single thermal coding estimated value of the frequency domain transmitting signal in the kth classification network unit in the constellation diagram is obtained by using the softmax functionThe expression of the softmax function is:
Wherein, C is the number of categories classified by the classification layer; the classification layer uses a loss function as a constraint, where the loss function includes a kl divergence function or a cross-entopy function, specifically as follows:
wherein, And P i is the single-heat coding label value of the frequency domain transmission signal in the constellation diagram under the ith category.
Further, in the deep learning network based on the classification idea, residual connection is added after the second full connection layer of the kth classification network unit, and the expression is as follows:
Where epsilon is the residual factor, Is the estimated value of the frequency domain transmitting signal in the k-1 classification network unit in the constellation diagram; residual connection is also added after the classification layer, and the expression is:
wherein, Is the one-hot encoded estimate of the frequency domain transmitted signal in the constellation in the k-1 th classified network element.
Further, in step S2, in order to reduce the calculation amount of the deep learning network based on the classification idea, the input data of the deep learning network based on the classification idea is subjected to block processing, so as to reduce the data amount sent into the computer at one time;
The time-varying frequency domain channel matrix H after the block processing, the initial estimated values S 1 of the frequency domain received signal Y and the frequency domain transmitted signal in the constellation diagram are expressed as follows:
Y=[Γ1 Γ2 … ΓJ]
S1=[Υ1 Υ2 … ΥJ]
where J represents the total number of sub-blocks, Represents the J-th sub-block of the time-varying frequency-domain channel matrix H, j=1 to J, Γ j represents the J-th sub-block of the frequency-domain received signal Y, Y j represents the J-th sub-block of the initial estimate S 1 of the frequency-domain transmitted signal in the constellation.
The estimated value S of the initial frequency domain transmission signal in the constellation diagram is the initial value of the estimated value of the frequency domain transmission signal in the constellation diagram in the 1 st classified network unitIs obtained by initializing based on Zero Forcing (ZF) equalization method. The final estimated value S of the frequency domain transmission signal in the constellation diagram is the estimated value/>, in the constellation diagram, of the frequency domain transmission signal in the Kth classification network element
Further, the initial estimated value S 1 of the frequency domain receiving signal Y, the time-varying frequency domain channel matrix H and the frequency domain transmitting signal after the block processing in the constellation diagram is input into a deep learning network based on the classification idea, and offline training is performed in a tensorflow architecture, so that the output of the deep learning network based on the classification idea approaches to the single-heat coding tag value P of the frequency domain transmitting signal S and the frequency domain transmitting signal in the constellation diagram, and a deep classification network model after training is completed is obtained.
In step S3, in the specific time-varying channel communication, the input data to be equalized is subjected to block processing, input into a deep learning network based on the classification idea, and a deep classification network model after training is invoked to obtain the final single-heat coding estimated value of the frequency domain transmitting signal in the constellation diagram.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) The network structure is simplified by using the classification idea, and lower SER performance is realized with lower algorithm complexity.
(2) Meanwhile, the communication theory model and the diversity of the communication data set are utilized, the network has better robustness to different actually measured channels, and can adapt to channels with different complexity degrees and achieve better performance.
(3) The method can adapt to OFDM signals with different modulation modes.
(4) Can adapt to OFDM signals with different lengths.
Drawings
FIG. 1 is a block diagram of a kth classification network element in a deep learning network based on classification ideas according to the present invention;
FIG. 2 is a diagram of a residual error structure in a deep learning network based on a classification idea of the present invention;
Fig. 3 is a graph of the performance of the invention based on QPSK modulation for an underwater acoustic channel.
FIG. 4 is a graph of the performance of the present invention for an underwater acoustic channel based on 16-QAM modulation;
Fig. 5 is a graph of the performance of the input 1024 subcarrier signal facing the underwater acoustic channel in accordance with the present invention.
Detailed Description
In the following description, technical solutions are set forth in connection with specific illustrations in order to provide a full understanding of the present application. This application may be carried out in a number of other ways than those herein set forth, and similar embodiments would be apparent to those of ordinary skill in the art without undue burden from the present disclosure.
Example 1:
in this embodiment, underwater acoustic channel equalization is taken as an embodiment, and a time-varying channel OFDM signal equalization method based on a deep classification network includes the following steps:
S1, given an actual measurement channel data set, constructing a deep learning network based on a classification idea, and respectively constructing input and output of the deep learning network based on the classification idea;
in this embodiment, the OFDM signal is set to 512 subcarriers, and the actual measurement channel data set uses the underwater sound channel data set NOF acquired in oslo isthmus, norway, and the added additive noise signal-to-noise ratio (Signal to noise ratio, SNR) ranges from 10 dB to 35dB. Selecting 120 channels with equal length from the underwater sound channel data set NOF The time-varying channel impulse response ir is obtained by means of linear interpolation.
Performing two-dimensional Fourier transform on the time-varying channel impulse response to obtain a time-varying frequency domain channel matrix H, wherein the time-varying frequency domain channel matrix H is specifically as follows:
the time-varying channel signal transmission process is modeled as:
Y=HS+Z;
Wherein the method comprises the steps of Is a frequency domain received signal obtained by subjecting the time domain received signal y (n) to FFT processing,Is a frequency domain transmission signal obtained by FFT processing of a time domain transmission signal s (n)/>Is the frequency domain additive noise obtained after the time domain additive noise z (n) is subjected to FFT processing.
In this embodiment, the frequency domain transmission signal is Quadrature Phase Shift Keying (QPSK) modulated, the frequency domain transmission signal after the QPSK modulation is subjected to one-time thermal coding, and the frequency domain transmission signal after the QPSK modulation is subjected to one-time thermal coding to be used as the one-time thermal coding tag value P of the frequency domain transmission signal in the constellation diagram.
In this embodiment, the deep learning network based on the classification idea includes 22 classification network units with the same structure stacked, where the kth classification network unit sequentially includes an input layer, a first full-connection layer, an activation function layer, a second full-connection layer, and a classification layer, k=1, …,22;
As shown in fig. 1, in the kth sorting network unit, the input layer receives the input frequency domain received signal Y, the time-varying frequency domain channel matrix H, and the estimated values of the frequency domain transmitted signals in the kth-1 sorting network unit in the constellation And outputting the spliced input matrix to the first full-connection layer, wherein the input matrix is expressed as:
Wherein λ k1 is a first step size parameter in the kth classified network element, λ k2 is a second step size parameter in the kth classified network element; when k=1, input into the 1 st classified network element Initializing based on Zero Forcing (ZF) equalization method to obtain and optimize network performance; since the time-varying frequency domain channel matrix H is not necessarily a non-singular matrix, the Zero-forcing (ZF) -based equalization expression is a pseudo-inverse form/>
In a first fully connected layer, training and updating network weight parameters, the first fully connected layer is expressed as:
Wherein d k k-th classified network element is the output of the first full-connection layer, w k is the weight matrix of the k-th classified network element, and n k is the bias term of the k-th classified network element;
In this embodiment, the activation function adopted in the activation function layer adopts the tanh function, and θ k(dk) is limited between-1 and 1, and 0 is used as the center, so that the deep learning network based on the classification thought converges faster, and the expression is:
The output of the activation function layer is that the value A k=θk(dk),θk of the output d k of the first full-connection layer, which is subjected to nonlinear processing by the activation function, represents the activation function used by the kth classification network unit;
Output A k of the second full-connection layer input activation function layer outputs the estimated value of the frequency domain transmitting signal in the constellation diagram in the kth classification network unit
In the classification layer, according to the output A k of the input activation function layer, a softmax function is used to obtain a single thermal coding estimated value Q i of a frequency domain transmitting signal in a k-th classification network unit in a constellation diagram, wherein the softmax function has the expression:
In this embodiment, the value of C is 4, and the softmax function makes the sum of the probabilities of the 4 output items of the multi-classification layer to be 1, and enhances 1 output item to be enhanced to a number similar to 1 in the 4 output items, and suppresses the other three output items to a number similar to 0, so as to highlight the information to be enhanced;
The classification layer uses a loss function as a constraint, in this embodiment, the loss function uses a cross-entopy function to evaluate the single-heat-encoded estimated value of the frequency-domain transmission signal in the constellation diagram The distribution matching error with the tag value P is specifically as follows:
As shown in fig. 2, in the deep learning network based on the classification idea, residual connection is added after the second full connection layer of the kth classification network unit, where the expression is as follows:
Wherein the value of the residual factor epsilon is 0.1. Residual connection is also added after the classification layer, and the expression is:
The residual connection can avoid gradient disappearance of the deep learning network based on the classification idea, and optimize the effect of the deep learning network based on the classification idea.
The residual connection can avoid gradient disappearance of the deep learning network based on the classification idea, and optimize the effect of the deep learning network based on the classification idea.
S2, performing online training on the constructed deep learning network based on the classification ideas to obtain a deep classification network model after training is completed;
in order to reduce the calculated amount of the deep learning network based on the classification idea, the input data of the deep learning network based on the classification idea is subjected to block processing, and the data amount fed into a computer at one time is reduced;
The time-varying frequency domain channel matrix H after the block processing, the initial estimated values S 1 of the frequency domain received signal Y and the frequency domain transmitted signal in the constellation diagram are expressed as follows:
Y=[Γ1 Γ2 … ΓJ]
S1=[Υ1 Υ2 … ΥJ]
where J represents the total number of sub-blocks, Represents the J-th sub-block of the time-varying frequency domain channel matrix H, j=1 to J, Γ j represents the J-th sub-block of the frequency domain received signal Y, Y j represents the J-th sub-block of the initial estimated value S 1 of the frequency domain transmitted signal in the constellation map, in this embodiment, the value of J is 64;
The estimated value S of the initial frequency domain transmission signal in the constellation diagram is the initial value of the estimated value of the frequency domain transmission signal in the constellation diagram in the 1 st classified network unit Is obtained by initializing based on Zero Forcing (ZF) equalization method. The final estimated value S of the frequency domain transmission signal in the constellation diagram is the estimated value/>, in the constellation diagram, of the frequency domain transmission signal in the Kth classification network element
The method comprises the steps of inputting an initial estimated value S 1 of a frequency domain receiving signal Y, a time-varying frequency domain channel matrix H and a frequency domain transmitting signal which are subjected to block processing in a constellation diagram into a deep learning network based on a classification idea, and performing offline training in a tensorflow architecture, so that the output of the deep learning network based on the classification idea approximates to a single-heat coding tag value P of the frequency domain transmitting signal S and the frequency domain transmitting signal in the constellation diagram, and obtaining a deep classification network model after training;
S3, in specific time-varying channel communication, the input data to be balanced are subjected to blocking processing, a deep learning network based on a classification idea is input, a trained deep classification network model is called, a single-heat coding estimated value of a frequency domain transmitting signal in a constellation diagram is obtained, and time-varying channel OFDM signal equalization in different modulation modes is realized.
In this embodiment, through simulation experiments, the performance of the embodiment is shown in fig. 3, and compared with MMSE equalization algorithm, DFE equalization algorithm and SCN equalizer based on deep regression network, deep classification network obtains SER performance superior to other algorithms.
Example 2:
In this embodiment, unlike embodiment 1, the frequency domain transmission signal is subjected to 16-QAM modulation, the frequency domain transmission signal after 16-QAM modulation is subjected to one-time thermal encoding, and the frequency domain transmission signal after 16-QAM modulation is subjected to one-time thermal encoding as a one-time thermal encoding tag value P i of the frequency domain transmission signal in the constellation diagram.
In the classification layer, according to the output A k of the input activation function layer, the single thermal coding estimated value of the frequency domain transmitting signal in the kth classification network unit in the constellation diagram is obtained by using the softmax functionThe expression of the softmax function is:
In this embodiment, the value of C is 16, and the softmax function makes the sum of the probability of the 16 output items of the multi-classification layer to be 1, and enhances 1 output item to be enhanced to a number similar to 1 in the 16 output items, and suppresses the other three output items to a number similar to 0, so as to highlight the information to be enhanced;
In this embodiment, through simulation experiments, the performance of the embodiment is shown in fig. 4, and compared with the MMSE equalization algorithm, the DFE equalization algorithm, the deep classification network obtains SER performance superior to other algorithms. The invention has the advantages of being adaptable to different modulation modes and has better robustness.
Example 3:
in this example, unlike example 1, the OFDM signal is set to 1024 subcarriers, and the actual measurement channel data set uses the underwater sound channel data set NOF acquired in oslo isthmus, norway, and the added additive noise signal-to-noise ratio (Signal to noise ratio, SNR) ranges from 10 to 35dB. Selecting 120 channels with equal length from the underwater sound channel data set NOF The time-varying channel impulse response ir is obtained by means of linear interpolation.
The time-varying frequency domain channel matrix H after the block processing, the initial estimated values S 1 of the frequency domain received signal Y and the frequency domain transmitted signal in the constellation diagram are expressed as follows:
Y=[Γ1 Γ2 … ΓJ]
S1=[Υ1 Υ2 … ΥJ]
where J represents the total number of sub-blocks, Represents the J-th sub-block of the time-varying frequency domain channel matrix H, j=1 to J, Γ j represents the J-th sub-block of the frequency domain received signal Y, Y j represents the J-th sub-block of the initial estimated value S 1 of the frequency domain transmitted signal in the constellation map, in this embodiment, the value of J is 128;
In this embodiment, through simulation experiments, the performance of the embodiment is shown in fig. 5, and the deep classification network obtains SER performance superior to other algorithms by comparing the DFE equalization algorithm with the SCN equalizer based on the deep regression network. The invention can adapt to OFDM signals with different lengths on the premise of not changing the network structure, and has better practical application.
Claims (7)
1. The time-varying channel OFDM signal equalization method based on deep classification network is characterized by comprising the following steps:
S1, given an actual measurement channel data set, constructing a deep learning network based on a classification idea, and respectively constructing input and output of the deep learning network based on the classification idea;
The deep learning network based on the classification idea comprises K classification network units with the same structure, wherein the kth classification network unit sequentially comprises an input layer, a first full-connection layer, an activation function layer, a second full-connection layer and a classification layer, and k=1, … and K;
In the kth classifying network unit, the input layer receives the input frequency domain received signal Y, the time-varying frequency domain channel matrix H and the estimated value of the frequency domain transmitted signal in the kth-1 classifying network unit in the constellation diagram And outputting the spliced input matrix to the first full-connection layer, wherein the input matrix is expressed as:
Wherein λ k1 is a first step size parameter in the kth classified network element, λ k2 is a second step size parameter in the kth classified network element; when k=1, the initial estimated value of the frequency domain transmission signal in the constellation in the 1 st classified network element is input Initializing based on Zero Forcing (ZF) equalization method to obtain and optimize network performance; since the time-varying frequency domain channel matrix H is not necessarily a non-singular matrix, the Zero-forcing (ZF) -based equalization expression is in pseudo-inverse form
In a first fully connected layer, training and updating network weight parameters, the first fully connected layer is expressed as:
Wherein d k k-th classified network element is the output of the first full-connection layer, w k is the weight matrix of the k-th classified network element, and n k is the bias term of the k-th classified network element;
The activation function adopted in the activation function layer comprises relu functions or tanh functions, and the expressions are respectively:
relu(dk)=max(dk,0)
The output of the activation function layer is that the value A k=θk(dk),θk of the output d k of the first full-connection layer, which is subjected to nonlinear processing by the activation function, represents the activation function used by the kth classification network unit;
Output A k of the second full-connection layer input activation function layer outputs the estimated value of the frequency domain transmitting signal in the constellation diagram in the kth classification network unit
In the classification layer, according to the output A k of the input activation function layer, the single thermal coding estimated value of the frequency domain transmitting signal in the kth classification network unit in the constellation diagram is obtained by using the softmax functionThe expression of the softmax function is:
Wherein, C is the number of categories classified by the classification layer; the classification layer uses a loss function as a constraint, where the loss function includes a kl divergence function or a cross-entopy function, specifically as follows:
wherein, For the single-heat coding estimated value of the frequency domain transmitting signal in the constellation diagram in the kth classification network unit under the ith category, P i is the single-heat coding label value of the frequency domain transmitting signal in the constellation diagram under the ith category;
the deep learning network based on the classification idea is also added with residual connection after the second full connection layer of the kth classification network unit, and the expression is as follows:
Where epsilon is the residual factor, Is the estimated value of the frequency domain transmitting signal in the k-1 classification network unit in the constellation diagram; residual connection is also added after the classification layer, and the expression is:
wherein, Is the single-heat coding estimated value of the frequency domain transmitting signal in the k-1 classification network unit in the constellation diagram;
S2, performing online training on the constructed deep learning network based on the classification thought to obtain a deep classification network model;
S3, in specific time-varying channel communication, the input data to be balanced are subjected to blocking processing, a deep learning network based on a classification idea is input, an available deep classification network model is called, a single-heat coding estimated value of a frequency domain transmitting signal in a constellation diagram is obtained, and time-varying channel OFDM signal equalization in different modulation modes is realized.
2. The method for equalizing an OFDM signal of a time-varying channel based on a deep packet network as claimed in claim 1, wherein in step S1, the OFDM signal has N subcarriers, S (m) represents a signal transmitted on an mth subcarrier, and m=1 to N;
Where N represents the nth sampling time, n=0,..n-1; s (N) represents a time domain transmission signal after inverse fast fourier transform (INVERSE FAST fourier transform, IFFT) transformation at an nth sampling time N-point;
The time domain received signal y (n) of the OFDM time domain transmission signal s (n) arriving at the receiving end through the l paths at the nth sampling time is expressed as:
Where h (n, L) is the channel of the path of the nth sampling instant L, L is the total number of signal transmission paths, and z (n) is the time domain additive noise of the nth sampling instant;
the measured channel data set is a large number of actually measured channel sets; selecting X channels with equal length from the measured channel data set Obtaining a time-varying channel impulse response by using a linear interpolation mode, wherein the expression is as follows:
Where ζ is a coefficient of linear interpolation, ir is a time-varying channel impulse response generated using the measured channel dataset, x represents the number of channels selected, Represents the X-th selected channel, where x=1, …, X.
3. The deep classification network-based time-varying channel OFDM signal equalization method of claim 2, wherein the time-varying channel impulse response ir is subjected to two-dimensional Fourier transform to obtain a time-varying frequency domain channel matrix H, and the method is specifically as follows:
the time-varying channel signal transmission process is modeled as:
Y=HS+Z;
wherein, Is a frequency domain received signal obtained by subjecting the time domain received signal y (n) to FFT processing,Is a frequency domain transmission signal obtained by FFT processing of a time domain transmission signal s (n)/>Is the frequency domain additive noise obtained after the time domain additive noise z (n) is subjected to FFT processing.
4. A deep-layer-classification-network-based time-varying channel OFDM signal equalization method according to claim 3, wherein the frequency-domain transmission signal S is Quadrature Phase Shift Keying (QPSK) modulated or 16 Quadrature amplitude modulated (16-Quadrature Amplitude Modulation, 16-QAM);
and performing single-heat coding on the frequency domain transmitting signals after QPSK or 16-QAM modulation, wherein the single-heat coding is respectively expressed as follows:
or=1,oi=1→[1,0,0,0]
or=-1,oi=1→[0,1,0,0]
or=1,oi=-1→[0,0,1,0]
or=-1,oi=-1→[0,0,0,1]
or=-3,oi=-3→[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=-1→[0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=3→[0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0]
or=-3,oi=1→[0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=-3→[0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=-1→[0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0]
or=-1,oi=3→[0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0]
or=-1,oi=1→[0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0]
or=3,oi=-3→[0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0]
or=3,oi=-1→[0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0]
or=3,oi=3→[0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0]
or=3,oi=1→[0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0]
or=1,oi=-3→[0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0]
or=1,oi=-1→[0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0]
or=1,oi=3→[0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0]
or=1,oi=1→[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1]
Where o r is the real part of the frequency domain transmitted signal after QPSK or 16-QAM modulation, and o i is the imaginary part of the frequency domain transmitted signal after QPSK or 16-QAM modulation; and performing single-heat coding on the frequency domain transmitting signal after QPSK or 16-QAM modulation to obtain a single-heat coded tag value P of the frequency domain transmitting signal in a constellation diagram.
5. The method for equalizing time-varying channel OFDM signals based on deep classification network according to any one of claims 1 to 4, wherein in step S2, in order to reduce the computation amount of the deep learning network based on classification ideas, the input data of the deep learning network based on classification ideas is subjected to block processing, so as to reduce the data amount fed into the computer at one time;
The time-varying frequency domain channel matrix H after the block processing, the initial estimated value of the frequency domain receiving signal Y and the frequency domain transmitting signal in the constellation diagram Expressed in the following form:
Y=[Γ1 Γ2 … ΓJ]
where J represents the total number of sub-blocks, Represents the jth sub-block of the time-varying frequency domain channel matrix H, j=1 to J, Γ j represents the jth sub-block of the frequency domain received signal Y, γ j represents the initial estimate/>, in the constellation, of the frequency domain transmitted signalIs the j-th sub-block of (c).
6. The method for equalizing a time-varying channel OFDM signal based on deep-layer classification network as claimed in claim 5, wherein the frequency-domain received signal Y, the time-varying frequency-domain channel matrix H and the initial estimated values of the frequency-domain transmitted signal in the constellation after the block processing are dividedAnd inputting a deep learning network based on a classification idea, performing offline training at tensorflow architecture, enabling the output of the deep learning network based on the classification idea to approach to a single-heat coding tag value P of a frequency domain transmitting signal S and the frequency domain transmitting signal in a constellation diagram, and obtaining a deep classification network model after training.
7. The method for equalizing time-varying channel OFDM signals based on deep-layer classification network as claimed in claim 6, wherein in step S3, in specific time-varying channel communication, the input data to be equalized is subjected to block processing, input into deep learning network based on classification ideas, and a trained deep-layer classification network model is invoked to obtain final single-heat coding estimated value of the frequency domain transmitting signal in constellation diagram.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210109366.7A CN114528925B (en) | 2022-01-28 | 2022-01-28 | Time-varying channel OFDM signal equalization method based on deep classification network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210109366.7A CN114528925B (en) | 2022-01-28 | 2022-01-28 | Time-varying channel OFDM signal equalization method based on deep classification network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114528925A CN114528925A (en) | 2022-05-24 |
CN114528925B true CN114528925B (en) | 2024-04-26 |
Family
ID=81622886
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210109366.7A Active CN114528925B (en) | 2022-01-28 | 2022-01-28 | Time-varying channel OFDM signal equalization method based on deep classification network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114528925B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2007110333A (en) * | 2005-10-12 | 2007-04-26 | Nippon Hoso Kyokai <Nhk> | Ofdm receiver for performing estimated correction of channel |
CN109067688A (en) * | 2018-07-09 | 2018-12-21 | 东南大学 | A kind of OFDM method of reseptance of data model double drive |
CN113014524A (en) * | 2021-03-03 | 2021-06-22 | 电子科技大学 | Digital signal modulation identification method based on deep learning |
-
2022
- 2022-01-28 CN CN202210109366.7A patent/CN114528925B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2007110333A (en) * | 2005-10-12 | 2007-04-26 | Nippon Hoso Kyokai <Nhk> | Ofdm receiver for performing estimated correction of channel |
CN109067688A (en) * | 2018-07-09 | 2018-12-21 | 东南大学 | A kind of OFDM method of reseptance of data model double drive |
CN113014524A (en) * | 2021-03-03 | 2021-06-22 | 电子科技大学 | Digital signal modulation identification method based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN114528925A (en) | 2022-05-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111404849B (en) | OFDM channel estimation and signal detection method based on deep learning | |
CN113676431B (en) | Model-driven MIMO-OFDM receiving method without cyclic prefix | |
CN113381951A (en) | MFTN joint channel estimation and equalization method under time-frequency conversion selective fading channel | |
CN111786923A (en) | Channel estimation method for time-frequency double-channel selection of orthogonal frequency division multiplexing system | |
CN108881080B (en) | OFDM anti-ICI detection method based on sliding window and deep learning | |
CN113852575A (en) | Iterative OTFS symbol detection method based on time domain channel equalization assistance | |
CN115664898A (en) | OFDM system channel estimation method and system based on complex convolution neural network | |
CN112215335A (en) | System detection method based on deep learning | |
Hua et al. | Signal detection in uplink pilot-assisted multi-user MIMO systems with deep learning | |
CN115250216A (en) | Underwater sound OFDM combined channel estimation and signal detection method based on deep learning | |
Wang et al. | Online LSTM-based channel estimation for HF MIMO SC-FDE system | |
CN114499601B (en) | Large-scale MIMO signal detection method based on deep learning | |
Ye et al. | Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems | |
Zhou et al. | Model-driven deep learning-based signal detector for CP-free MIMO-OFDM systems | |
CN112636855A (en) | OFDM signal detection method | |
CN109818891B (en) | Lattice reduction assisted low-complexity greedy sphere decoding detection method | |
Liu et al. | RecNet: Deep learning-based OFDM receiver with semi-blind channel estimation | |
Belkacem et al. | Neural network equalization for frequency selective nonlinear MIMO channels | |
CN114528925B (en) | Time-varying channel OFDM signal equalization method based on deep classification network | |
Baumgartner et al. | A soft interference cancellation inspired neural network for SC-FDE | |
CN114006797A (en) | Multi-antenna equalization receiving method for high-speed video communication | |
CN111769975A (en) | MIMO system signal detection method and system | |
CN112039809A (en) | Block iterative equalizer based on mixed soft information and bidirectional block iterative equalizer | |
Ajayi et al. | Low-complexity neural networks for denoising imperfect CSI in physical layer security | |
Hanoon et al. | Performance Evaluation of SISO-OFDM Channel Equalization Utilizing Deep Learning |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |