CN112637093B - Signal detection method based on model-driven deep learning - Google Patents
Signal detection method based on model-driven deep learning Download PDFInfo
- Publication number
- CN112637093B CN112637093B CN202011451458.0A CN202011451458A CN112637093B CN 112637093 B CN112637093 B CN 112637093B CN 202011451458 A CN202011451458 A CN 202011451458A CN 112637093 B CN112637093 B CN 112637093B
- Authority
- CN
- China
- Prior art keywords
- network
- model
- data
- channel
- dnn
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 39
- 238000013135 deep learning Methods 0.000 title claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 34
- 238000012545 processing Methods 0.000 claims abstract description 7
- 238000004891 communication Methods 0.000 claims description 34
- 230000006870 function Effects 0.000 claims description 17
- 238000000034 method Methods 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 13
- 230000004044 response Effects 0.000 claims description 11
- 210000002569 neuron Anatomy 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 230000009467 reduction Effects 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 4
- 238000009499 grossing Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 2
- 238000003780 insertion Methods 0.000 claims description 2
- 230000037431 insertion Effects 0.000 claims description 2
- 238000013507 mapping Methods 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims 1
- 238000010606 normalization Methods 0.000 claims 1
- 238000003062 neural network model Methods 0.000 abstract description 6
- 238000005516 engineering process Methods 0.000 description 6
- 238000013461 design Methods 0.000 description 4
- 210000004027 cell Anatomy 0.000 description 3
- 239000013256 coordination polymer Substances 0.000 description 3
- 125000004122 cyclic group Chemical group 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000005562 fading Methods 0.000 description 1
- 238000012966 insertion method Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 239000003016 pheromone Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- 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
-
- 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
- G06N3/048—Activation functions
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- H04L25/0202—Channel estimation
- H04L25/0224—Channel estimation using sounding signals
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/26—Systems using multi-frequency codes
- H04L27/2601—Multicarrier modulation systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/26—Systems using multi-frequency codes
- H04L27/2601—Multicarrier modulation systems
- H04L27/2647—Arrangements specific to the receiver only
- H04L27/2655—Synchronisation arrangements
- H04L27/2689—Link with other circuits, i.e. special connections between synchronisation arrangements and other circuits for achieving synchronisation
- H04L27/2695—Link with other circuits, i.e. special connections between synchronisation arrangements and other circuits for achieving synchronisation with channel estimation, e.g. determination of delay spread, derivative or peak tracking
-
- 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
Abstract
The invention relates to a signal detection method based on model-driven deep learning. The invention establishes a channel estimation and signal detection model based on an OFDM system. The channel estimation adopts a combined neural network model taking an MMSE estimator based on DFT and FC-DNN as sub-networks, pilot frequency data which are distributed in a self-adaptive mode are preprocessed through the MMSE estimator, DNN network initialization information is extracted, and a more accurate channel estimation model is obtained according to a training learning network ChannelEstNe. SignalDetNet adopts a ZF equilibrium detection preprocessor, LSTM and DNN to form a combined network, realizes final signal detection and restores the original signal. The structure keeps the form of processing signals block by block of the OFDM system, can restore the transmitted data in the OFDM system with linear and nonlinear distortion, and has higher initial training speed by combining with the traditional algorithm, thereby improving the deployment efficiency.
Description
Technical Field
The invention belongs to the field of intelligent communication, relates to a signal detection method based on model-driven deep learning, and particularly relates to a signal detection method of an OFDM wireless communication receiver based on model driving and deep learning.
Background
The OFDM receiver scheme mainly comprises two functional modules of channel estimation and signal detection, namely accurate Channel State Information (CSI) is obtained through channel estimation at first, and then the CSI obtained through estimation is used for recovering a sending signal. Most of the traditional channel estimation and signal detection technologies improve the receiving performance of a communication system through a complex algorithm, but for 5G wireless communication which requires high dimensionality, high speed and high density at present, the effectiveness of communication is greatly influenced by higher complexity calculation. The intelligent communication applies artificial intelligence to each layer of wireless communication, realizes the organic integration of artificial technology and the traditional communication system, and combines deep learning with the wireless communication system to greatly improve the system efficiency.
The artificial intelligence applied to the research of the physical layer of the wireless communication system mainly comprises two types of deep learning networks, one type is based on data driving, and the other type is based on model driving. The deep learning network based on the model replaces a certain module or training related parameters to improve the performance of the module on the basis of the original model of the communication system, and compared with the deep learning network based on data driving, the deep learning network based on the model can obviously reduce the information amount of training and has better generalization and adaptivity.
Disclosure of Invention
The invention solves the problem of low signal reliability caused by the traditional channel estimation and signal detection algorithm in the existing Orthogonal Frequency Division Multiplexing (OFDM) receiver, and provides an algorithm for realizing channel estimation and signal detection of received signals by respectively adopting ChannelEstNet and SignalDetNet neural network models.
The invention adopts the technical scheme that an OFDM signal detection receiver based on model-driven deep learning comprises the following steps:
step 1: the data sets required for the channeliestnet and SignalDetNet are generated based on the OFDM wireless communication system framework. The characteristic information of the ChannelEstNet data set comes from pilot frequency information of a sending signal and pilot frequency information of a receiving signal, and a label is distributed according to a channel response matrix. The characteristic information of the SignalDetNet data set is from channel estimation information and received signal information output by ChannelEstNet, and labels are distributed according to original sending symbols.
Step 2: randomly disordering the data set samples of the two models in the step 1 and reclassifying the data set samples into 12000 training sets for offline training, 3000 verification sets for verifying the performance of the models and 2000 test sets for online testing of the network performance;
and step 3: based on the sample data in the step 2, linear dimension reduction is carried out on the characteristic information by utilizing a Linear Discriminant Analysis (LDA) algorithm, and the characteristic information of various dimensions is used as the input of a ChannelEstNet and SignalDetNet network model;
and 4, step 4: the channel estimation module ChannelEstNet and the signal detection module SignalDetNet are both constructed by combining a deep learning network and a communication algorithm in a block-by-block signal processing mode similar to a conventional communication system, wherein the initialization of a ChannelEstNet sub-network DNN-1 is weighted by MMSE and LS algorithms, and the sub-network input of the SignalDetNet adopts ZF pre-detection processed data.
And 5: based on the characteristic information of the sample data set generated in step 3, the ChannelEstNet performs DNN-1 network training to estimate the channel frequency domain response at the OFDM symbol by using the channel frequency domain response at the pilot generated by the received pilot and the known transmitted pilot input MMSE preprocessing, performs time domain smoothing by using DFT technique, performs DNN-2 network training again to obtain channel estimation information, then performs ZF detection on the channel estimation information and the received data symbol to generate xZF of preliminary estimation, and inputs the sinaldetnet sub-network LSTM + DNN, that is, the network combining LSTM and DNN to perform further signal detection training.
Step 6: and (3) carrying out online test according to the model established in the step (1), the step (2), the step (3), the step (4) and the step (5) to replace a receiving end channel estimation and signal detection part of the OFDM system.
Preferably, in step 1, the data sets of the networks channeltnet and SignalDetNet are generated based on the OFDM system for the channel conditions of different environments indoors and outdoors, respectively, the pilot is inserted in the OFDM wireless communication system by using an adaptive pilot allocation method, the data set of channeltnet uses the pilot and the Channel State Information (CSI) at the OFDM symbol, the label is the Channel State Information (CSI) at the OFDM symbol, and the label of SignalDetNet is set as the transmitted data symbol.
The adaptive pilot frequency insertion method in step 1 is as follows:
(1) selecting the K frame data to be transmitted by adopting the minimum proportion of pilot frequency occupying subcarriers when K is 1;
(2) determining the pilot frequency-data ratio MK adopted by the next frame according to the estimated channel parameters obtained from the Kth frame data at the receiving end;
(3) k is K +1, the transmitting end inserts pilot frequency in the Kth frame data according to the pilot frequency and data ratio MK;
(4) and (3) skipping to the step (2) for estimating the channel parameters and judging the pilot frequency occupation ratio of the next frame.
Preferably, in step 2, the data set, especially the channel parameters, are normalized, and for the convenient complex form data input in the training process, the channel data are independently divided and stored in the real part and the imaginary part.
Preferably, in step 3, the original feature information and the feature information after the dimensionality reduction of the LDA are used as two types of feature information, and the network model of the receiving end optimizes parameters according to the two types of feature information to increase the expansibility of the model.
Preferably, in the step 4, the channeliEstNet subnetwork DNN-1 has 5 layers, wherein 3 layers are hidden layers, the number of neurons in each layer is 16, 1024 and 128, the number of neurons in 3 layers of DNN-2 is 128,1024,128 respectively, and the activation function is Tanh; the SignalDetNet sub-network LSTM adopts 3 layers of hidden layers, the time step is set to be 64, 30, 20 and 16 hidden units are respectively arranged, the state activation function is a Relu function, the sub-network DNN-2 has 3 layers in total, each layer respectively contains 64, 64 and 48 neurons, the activation function is a Sigmoid function, and the optimization schemes of the two sub-networks adopt a mean square error cost function and adaptive moment estimation (Adam).
The 48 output layer neurons in step 4 correspond to 48 bits estimated from 8 consecutive subcarriers, e.g. 64-QAM, with 6 bits per symbol. Because the logical sigmoid function maps the input to the interval [0, 1], if the output is greater than 0.5, the received binary symbol will be "1", otherwise "0".
Preferably, in step 5, in the ChannelEstNet, MMSE calculates (channel frequency domain response) CFR at the pilot frequency through characteristic parameters, then CFR at each OFDM symbol position is generated through a sub-network DNN-1, time domain smoothing is performed through a DFT-based channel estimation method, and then more accurate (channel state information) CSI is obtained through a sub-network DNN-2. The computational parameters of SignalDetNet include the received data signal and the data channel response generated by ChannelEstNet, and the subnetwork LSTM + DNN combines ZF to preprocess the detected input data xZF to generate the original modulation signal.
In step 5, the training process of the ChannelEstNet and SignalDetNet includes the following steps:
step 5.1: 12000 groups of sample data are input into ChannelEstNet for training, the optimal weight and the bias value are obtained by adopting the mean square error cost function optimization, 2000 periods (Epochs) are trained, the learning rate is set as a step function, the initial value is set as 0.001, and the learning rate is reduced by 10 times in every 1000 periods.
Step 5.2: the ChannelEstNet model is fixed to train SignalDetNet, an adaptive moment estimation optimizer (Adam) is adopted to obtain the optimal weight and bias, 5000 periods are trained, the initial value of the learning rate is set to be 0.001, and each 2000 period is reduced by 5 times.
Step 5.3: and verifying the established network performance by using 3000 groups of verification set characteristic information.
Compared with the prior art, the signal detection method of the OFDM wireless communication receiver based on the model-driven deep learning has the following technical effects by adopting the technical scheme:
the invention relates to a signal detection method of an OFDM wireless communication system based on model-driven deep learning, which innovatively adopts a deep learning network to be combined with the traditional communication professional technology, designs a receiving end model structure based on OFDM design, and respectively replaces a channel estimation part and a signal detection part of the system with ChannelEstNet and SignalDetNet. The ChannelEstNet is a neural network model with a combination network of MMSE channel estimators and a sub-network, and the model realizes channel estimation by a pilot-based auxiliary channel estimation method. The SignalDetNet is a neural network model taking a ZF receiver and an LSTM + FC-DNN combined network as a sub-network, performs signal detection according to network channel estimation parameters of the ZF receiver and the LSTM + FC-DNN combined network, and can restore transmitted data in an OFDM system with linear and nonlinear distortion. The two networks keep the block-by-block construction of a conventional communication system, the original data is preprocessed by adopting the conventional technology, and then the initialization of the networks is processed by adopting the conventional algorithm through the networks, so that the training process is greatly accelerated, and the deployment speed is increased. For different communication systems, the method only needs to generate a data set according to the used system framework, has stronger generalization and is more suitable for the requirements of the current 5G communication.
Drawings
FIG. 1 is a block diagram of a model driven OFDM wireless communication system
FIG. 2 Structure of ChannelEstNet
FIG. 3 is a diagram of a SignalDetNet structure
FIG. 4DNN-1 model Structure
FIG. 5DNN-2 model Structure
FIG. 6LSTM network architecture
FIG. 7 model-driven OFDM wireless communication system error rate performance
Detailed Description
The invention will be further described with reference to specific embodiments, and the advantages and features of the invention will become apparent as the description proceeds. The examples are illustrative only and do not limit the scope of the present invention in any way. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention, and that such changes and substitutions are intended to be within the scope of the invention.
Step 1: the data sets required for the deep learning model are generated based on the OFDM wireless communication system framework. The characteristic information of the data set comes from the characteristics of the received signals at the receiving end, and the CSI, the channel state information of the transmitted data symbols and the originally sent data symbols are respectively the training labels of the channel estimation and signal detection models aiming at different indoor and outdoor channel conditions.
In the step, the specific implementation process of the OFDM wireless communication system includes the following steps:
step 1.1: bit information to be transmitted at a transmitting end is converted into a plurality of sub-bit data streams through serial-parallel conversion, then each path of bit data is digitally modulated, the modulated data is loaded on each subcarrier in an IDFT mode, and then the modulated data is transmitted through an antenna after being subjected to parallel-serial conversion and insertion of a guard interval (CP). And N subcarriers are shared in one OFDM symbol, and the pilot frequency is inserted into the OFDM symbol by adopting a self-adaptive pilot frequency distribution method.
Step 1.2: through the frequency domain selective fading multipath channel, the received signal can be represented as:
wherein
Representing a cyclic convolution, x (n) and w (n) representing the transmitted signal and noise. Removing the CP and performing DFT, wherein the channel frequency domain response of the K sub-carrier is as follows:
where g (n) is the channel impulse response. In one OFDM symbol time, the channel response vector on each pilot subcarrier is: h ═ h (1), …, h (N)P)]TSo that y is Xh + z, y being the received symbol vector on the pilot subcarriers of one OFDM symbol time; z represents additive white gaussian noise.
Step 1.3: after the CP is removed at the receiving end, the system can convert the linear convolution of the time domain signal and the channel impulse response into the cyclic convolution, and simultaneously can convert the cyclic convolution on the time domain into the frequency domain, which is expressed as the property of point-by-point multiplication of the frequency domain signal and the corresponding frequency domain subchannel gain. Therefore, on the kth frequency domain subchannel, the received signal is y (k) ═ x (k) h (k) + w (k).
Step 2: randomly disordering the sample data in the step 1, dividing the sample data into 12000 training sets, 3000 testing sets and 2000 testing sets again, and normalizing the data to enable the data range to be (0, 1).
And step 3: and (3) performing dimension reduction processing on the characteristic information by utilizing a Linear Discriminant Analysis (LDA) algorithm based on the training data in the step (2). The implementation of the LDA algorithm comprises the following steps:
step 3.1: calculating a mean vector μ for each class sample in the datasetjAnd an overall mean vector μ.
Step 3.2: calculating an internal divergence matrix SwPowder of general medicineDegree matrix StThen obtain the inter-class divergence matrix Sb＝St-Sw。
Step 3.3: for matrix
SbAnd (4) decomposing the characteristic values and arranging the characteristic values from large to small. 4
Step 3.4: arranging the first n eigenvectors w from big to small of the eigenvalue1,w2,…,wnThe N-dimensional samples are mapped to the N-dimensions by the following mapping:
and 4, step 4: based on the data sets in the step 2 and the step 3, channel estimation and signal detection are respectively carried out by utilizing the ChannelEstNet and the SignalDetNet, the structure keeps the conventional form of block-by-block processing of the OFDM wireless communication system, and a neural network is constructed by combining the traditional communication signal technology, so that the rapid convergence of training time is accelerated, and the performance of the OFDM receiver is effectively improved.
As shown in fig. 2, the ChannelEstNet algorithm is implemented by a DFT-based MMSE estimator in combination with a DNN network. The channel frequency domain response at pilot input DNN-1 sub-network is pre-processed by the MMSE approach, whose output can be expressed as:
wherein
Which represents the estimate of the LS channel and,
and RHHAre the true channel vectors H and H, respectively, in the frequency domain
And H's autocorrelation matrix.
Then DNN-1 (see fig. 4) to generate a preliminary channel estimate
Performing DFT time domain smoothing to eliminate noise beyond maximum time delay
Finally, it is input into the signed channel state information obtained by training DNN-2 (see FIG. 5).
As shown in FIG. 3, the SignalDetNet network model structure is composed of three layers of LSTM with 30, 20 and 16 hidden units for each layer of sub-network LSTM and three layers of FC-DNN with 64, 64 and 48 neurons for each layer of FC-DNN network. The LSTM neural network is optimized through an improved ant colony optimization algorithm, the ant colony optimization algorithm determines the intensity of pheromones to change the search randomness according to the comparison condition of data to be predicted and model training data through an adaptive factor P self-adaptive adjustment information enlightenment factor alpha.
As shown in FIG. 6, each LSTM unit completes long-term memory of information and further extraction of feature information through cooperation of a forgetting gate, an input gate and an output gate, wherein the forgetting gate is used for outputting H through the LSTM unit at the previous momentt-1And sample input x at the current timetSplicing into a new input variable, and outputting a value between 0 and 1 through calculation of a full connection layer: f. oft＝σ(Wf[ht-1,xt]+bf). The output gate adds the previous state information and the current new information to obtain a new cell state memory, wherein the forgetting gate and the input gate are used for respectively controlling the degree of the two parts of information to be kept, and the new cell state: ct＝ft*Ct-1+it*C* tThe output gate controls whether the LSTM unit is to output the current cell state information Ct，ot＝σ(Wo[ht-1,xt]+bo)，ht＝ot*tanh(Ct). After t time steps, the output of the LSTM can be obtained as:
FC-DNN conforms the output dimension of the network to the class of modulation scheme, e.g., 48-QAM, with 48 output neurons corresponding to 48 bits to be estimated from eight consecutive subcarriers (6 bits per symbol), and the formula is:
in order to accelerate the convergence of training rapidly by the two model-driven neural networks, network weights are initially set. Initializing DNN-1 network in ChannelEstNet by real-valued Linear Minimum Mean Square Error (LMMSE) weight matrix:
and 5: and (3) respectively replacing the modulation and modulation identification parts of the OFDM system according to the models established in the steps 1, 2, 3 and 4 to carry out online test, and obtaining the error rate performance of the system, as shown in figure 7.
The design of the technical scheme is based on a signal detection method of an OFDM wireless communication system receiver based on model-driven deep learning, a deep learning network is combined with the traditional communication professional technology, a receiving end model structure based on OFDM design is designed, and a channel EstNet and a SignalDetNet respectively replace a channel estimation part and a signal detection part of the system.
ChannelEstNet is a neural network model employing a combination of a DFT-based MMSE estimator and DNN. The characteristic information extracted from each OFDM transmission symbol is input into a sub-network to be used for the next signal detection according to the optimal channel state parameter obtained by training. The SignalDetNet forms a combined neural network model by a ZF equilibrium detection preprocessor, an LSTM and an FC-DNN. The scheme can effectively solve the problem that the transmitted data in the OFDM system with linear and nonlinear distortion can be recovered, and the neural network training speed is higher, thereby meeting the requirements of the current 5G communication.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.
Claims (6)
1. A signal detection method based on model-driven deep learning is characterized by comprising the following steps:
step 1: generating a data set required by ChannelEstNet and SignalDetNet based on an OFDM wireless communication system framework, wherein the characteristic information of the ChannelEstNet data set is from a channel frequency domain at a pilot frequency position of a sending signal and pilot frequency information of a receiving signal, a channel response matrix distribution label at a root symbol, and the characteristic information of the SignalDetNet data set is from channel estimation information and receiving signal information output by ChannelEstNet, and the label is distributed according to an original sending symbol;
step 2: normalizing the data set samples of the ChannelEstNet and the SignalDetNet in the step 1, and subdividing the data set samples into 12000 groups of training sets for offline training, 3000 groups of verification sets for verifying the performance of the model and 2000 groups of test sets for online testing of the network performance;
and step 3: based on the training sample data in the step 2, linear dimension reduction is carried out on the feature information by utilizing a Linear Discriminant Analysis (LDA) algorithm, and the feature information of various dimensions is used as the input of a ChannelEstNet and SignalDetNet network model;
and 4, step 4: inputting pilot information in the received signals in the data in the step 3 and pilot information of a sending end into a ChannelEstNet network for self-learning, inputting the received signal information and channel information CSI estimated by the ChannelEstNet network into a SignalDetNet network for self-learning, wherein both the deep learning network and a communication algorithm are combined and constructed in a block-by-block signal processing mode similar to a conventional communication system, the initialization of a ChannelEstNet sub-network DNN-1 is set by LMMSE and LS, and the SignalDetNet network adopts ZF pre-detection processing on LSTM data to be processed; a channel estimation module: the ChannelEstNet adopts a model-driven deep learning construction method and is formed by combining a DNN network with a DFT-based MMSE estimator, wherein the DNN-1 sub-network has 5 layers, 3 layers of hidden layers are provided, and the number of neurons in each layer is 16, 1024 and 128; the number of sub-network DNN-2 in 3 layers is 128,1024 and 128, and the activation functions are Tanh; the signal detection module: the SignalDetNet is constructed by combining a ZF detector, an LSTM network and an FC-DNN network, wherein the LSTM sub-network is provided with 3 layers of hidden layers, the time step is set to be 64, the LSTM sub-network is respectively provided with 30, 20 and 16 hidden units, and the state activation function is a Relu function; the FC-DNN sub-network has 3 layers of 64 neurons, 64 neurons and 48 neurons respectively, and the activation function adopts a Sigmoid function;
and 5: the channel EstNet adopts a channel information parameter h at the pilot frequency generated by MMSE preprocessing of a received pilot frequency and a known transmitted pilot frequency inputMMSEDNN-1 training is carried out to generate preliminarily estimated channel frequency domain state information at OFDM symbols, IDFT time domain smoothing denoising is carried out on the channel frequency domain state parameters, DNN-2 training is carried out after DFT is converted to a frequency domain to obtain final channel state information, and then ZF detection processing is carried out on the channel estimation information and received data symbols to generate preliminarily estimated modulated signal XZFInputting the estimation information and the received data signal into a SignalDetNet sub-network LSTM + DNN for signal detection model training;
step 6: and (3) carrying out online test according to the model established in the step (1), the step (2), the step (3), the step (4) and the step (5) to replace a receiving end channel estimation and signal detection part of the OFDM system.
2. The signal detection method based on model-driven deep learning according to claim 1, characterized in that: in the step 2, the data set is processed by normalization, the training mode is that the data is input into the neural network in a complex form, and the data of the channel is used for independently splitting and storing the real part and the imaginary part.
3. The signal detection method based on model-driven deep learning according to claim 1, characterized in that: the pilot frequency insertion in the OFDM wireless communication system adopts a self-adaptive pilot frequency distribution method as follows:
(1) selecting the K frame data to be transmitted by adopting the minimum proportion of pilot frequency occupying subcarriers when K is 1;
(2) determining the pilot frequency and data ratio M adopted by the next frame according to the estimated channel parameters obtained from the Kth frame data at the receiving endK；
(3) K is K +1, the transmitting end calculates the ratio M according to the pilot frequency and the dataKInserting pilot frequency in the Kth frame data;
(4) and (3) skipping to the step (2) for estimating the channel parameters and judging the pilot frequency occupation ratio of the next frame.
4. The signal detection method based on model-driven deep learning according to claim 1, characterized in that: the linear discriminant analysis LDA algorithm in the step 3 for performing linear dimensionality reduction on the feature information comprises the following steps:
step 3.1: calculating a mean vector μ for each class sample in the datasetjAnd an overall mean vector μ;
step 3.2: calculating an internal divergence matrix SwGlobal divergence matrix StThen obtain the inter-class divergence matrix Sb＝St-Sw；
Step 3.3: for matrix
Carrying out characteristic value decomposition and arranging the characteristic values from large to small;
step 3.4: arranging the first n eigenvectors w from big to small of the eigenvalue1,w2,…,wnThe N-dimensional samples are mapped to the N-dimensions by the following mapping:
5. the signal detection method based on model-driven deep learning according to claim 1, characterized in that: in step 5, the training process of the ChannelEstNet and SignalDetNet includes the following steps:
step 5.1: inputting 12000 groups of sample data into ChannelEstNet training, optimizing by adopting a mean square error cost function to obtain an optimal weight and an optimal offset value, training 2000 periods, setting a learning rate as a step function, setting an initial value as 0.001, and reducing the learning rate by 10 times in each 1000 periods;
step 5.2: fixing a ChannelEstNet model to train SignalDetNet, adopting an adaptive moment estimation optimizer (Adam) to obtain an optimal weight and bias, training 5000 periods, setting an initial value of a learning rate to be 0.001, and reducing 5 times in every 2000 periods;
step 5.3: the network optimization method adopts an MSE loss function and an Adam optimizer;
step 5.4: and verifying the performance of the established network model by using 3000 groups of verification set characteristic information.
6. The signal detection method based on model-driven deep learning according to claim 5, wherein: the method for the channel EstNet training comprises the following steps: initializing a ChannelEstNet model sub-network DNN-1, and setting a weight matrix through a linear minimum mean square error LMMSE algorithm
Wherein
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011451458.0A CN112637093B (en) | 2020-12-09 | 2020-12-09 | Signal detection method based on model-driven deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011451458.0A CN112637093B (en) | 2020-12-09 | 2020-12-09 | Signal detection method based on model-driven deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112637093A CN112637093A (en) | 2021-04-09 |
| CN112637093B true CN112637093B (en) | 2022-05-17 |
Family
ID=75309742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011451458.0A Active CN112637093B (en) | 2020-12-09 | 2020-12-09 | Signal detection method based on model-driven deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112637093B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113286309B (en) * | 2021-05-18 | 2023-02-07 | 合肥工业大学 | Heterogeneous communication method and system based on CSI |
| CN113285902B (en) * | 2021-05-19 | 2023-03-14 | 南京航空航天大学 | Design method of OFDM system detector |
| CN113676431B (en) * | 2021-07-08 | 2023-01-10 | 东南大学 | Model-driven MIMO-OFDM receiving method without cyclic prefix |
| CN114006794B (en) * | 2021-10-09 | 2022-11-25 | 苏州大学 | Complex value neural network-based channel estimation method and system |
| CN113872911B (en) * | 2021-10-15 | 2023-10-24 | 齐鲁工业大学 | Model-driven method and system for suppressing peak-to-average ratio of orthogonal frequency division multiplexing system |
| CN114584448B (en) * | 2022-02-16 | 2023-09-22 | 山东大学 | SM-OFDM signal grouping detection method based on deep neural network |
| CN114696933B (en) * | 2022-04-01 | 2023-02-07 | 西安交通大学 | AI receiver based on deep learning technology and use method |
| CN114759997B (en) * | 2022-04-08 | 2023-06-20 | 山东大学 | MIMO system signal detection method based on data model double driving |
| CN115250216A (en) * | 2022-07-19 | 2022-10-28 | 西安科技大学 | Underwater sound OFDM combined channel estimation and signal detection method based on deep learning |
| CN115356694B (en) * | 2022-08-26 | 2023-08-22 | 哈尔滨工业大学（威海） | High-frequency ground wave radar anti-impact interference method, system, computer equipment and application |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2007137484A1 (en) * | 2006-05-11 | 2007-12-06 | Shanghai Jiao Tong University | A channel estimation method and the device thereof |
| CN111224905A (en) * | 2019-12-25 | 2020-06-02 | 西安交通大学 | Multi-user detection method based on convolution residual error network in large-scale Internet of things |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090103666A1 (en) * | 2006-04-03 | 2009-04-23 | National Ict Australia Limited | Channel estimation for rapid dispersive fading channels |
| US7978776B2 (en) * | 2008-01-16 | 2011-07-12 | Southern Taiwan University Of Technology | Channel estimation device and related method of an orthogonal frequency division multiplexing system |
| CN107026804A (en) * | 2017-05-02 | 2017-08-08 | 南京航空航天大学 | Channel estimation methods based on exponential smoothing in MIMO ofdm systems |
| CN109067688B (en) * | 2018-07-09 | 2021-09-07 | 东南大学 | Dual-drive OFDM (orthogonal frequency division multiplexing) receiving method of data model |
| US20200153535A1 (en) * | 2018-11-09 | 2020-05-14 | Bluecom Systems and Consulting LLC | Reinforcement learning based cognitive anti-jamming communications system and method |
| CN109995449B (en) * | 2019-03-15 | 2020-12-18 | 北京邮电大学 | Millimeter wave signal detection method based on deep learning |
| CN111404849B (en) * | 2020-03-20 | 2021-01-12 | 北京航空航天大学 | OFDM channel estimation and signal detection method based on deep learning |
| CN111614587B (en) * | 2020-05-25 | 2021-04-06 | 齐鲁工业大学 | SC-FDE system signal detection method based on self-adaptive integrated deep learning model |
-
2020
- 2020-12-09 CN CN202011451458.0A patent/CN112637093B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2007137484A1 (en) * | 2006-05-11 | 2007-12-06 | Shanghai Jiao Tong University | A channel estimation method and the device thereof |
| CN111224905A (en) * | 2019-12-25 | 2020-06-02 | 西安交通大学 | Multi-user detection method based on convolution residual error network in large-scale Internet of things |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112637093A (en) | 2021-04-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112637093B (en) | Signal detection method based on model-driven deep learning | |
| Ali et al. | Automatic modulation classification of digital modulation signals with stacked autoencoders | |
| CN109067688B (en) | Dual-drive OFDM (orthogonal frequency division multiplexing) receiving method of data model | |
| CN109246038B (en) | Dual-drive GFDM receiver and method for data model | |
| CN111614587B (en) | SC-FDE system signal detection method based on self-adaptive integrated deep learning model | |
| Tian et al. | Modulation-constrained clustering approach to blind modulation classification for MIMO systems | |
| CN110113288B (en) | Design and demodulation method of OFDM demodulator based on machine learning | |
| Mao et al. | RoemNet: Robust meta learning based channel estimation in OFDM systems | |
| CN112600772B (en) | OFDM channel estimation and signal detection method based on data-driven neural network | |
| CN112235023B (en) | MIMO-SCFDE self-adaptive transmission method based on model-driven deep learning | |
| Ha et al. | Signal detection scheme based on adaptive ensemble deep learning model | |
| CN111669344B (en) | Time-varying OFDM system signal detection method based on deep learning | |
| CN109474388B (en) | Low-complexity MIMO-NOMA system signal detection method based on improved gradient projection method | |
| CN109787926A (en) | A kind of digital signal modulation mode recognition methods | |
| Liao et al. | A novel physical layer authentication method with convolutional neural network | |
| CN112953862A (en) | High-speed railway scene channel estimation method based on deep learning | |
| KR101979394B1 (en) | Adaptive transmission scheme determination apparatus based on MIMO-OFDM System using machine learning model and adaptive transmission method the same | |
| CN114513394B (en) | Signal modulation format identification method, system and device based on attention mechanism diagram neural network and storage medium | |
| CN112215335A (en) | System detection method based on deep learning | |
| Ye et al. | Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems | |
| CN113285902B (en) | Design method of OFDM system detector | |
| KR102064301B1 (en) | Signal detection apparatus using ensemble machine learning based on MIMO system and method thereof | |
| Zhang et al. | Efficient residual shrinkage CNN denoiser design for intelligent signal processing: Modulation recognition, detection, and decoding | |
| Li et al. | Model-driven deep learning scheme for adaptive transmission in MIMO-SCFDE system | |
| Yıldırım et al. | Deep receiver design for multi-carrier waveforms using cnns |
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 | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Li Jun Inventor after: Li Wenxin Inventor after: He Bo Inventor after: Fu Wen Wen Inventor after: Zhang Shaowei Inventor after: Han Yongli Inventor after: Shi Jun Inventor after: Gao He Inventor before: Li Jun Inventor before: Fu Wen Wen Inventor before: Li Wenxin Inventor before: Zhang Shaowei Inventor before: Han Yongli Inventor before: Shi Jun |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP01 | Change in the name or title of a patent holder | ||
| CP01 | Change in the name or title of a patent holder |
Address after: 250353 University Road, Changqing District, Ji'nan, Shandong Province, No. 3501 Patentee after: Qilu University of Technology (Shandong Academy of Sciences) Address before: 250353 University Road, Changqing District, Ji'nan, Shandong Province, No. 3501 Patentee before: Qilu University of Technology |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20231114 Address after: 264200 room a0216, 0218, Chuangye building, Shenyang road-108-2, Yiyuan sub district office, gaoqu District, Weihai City, Shandong Province Patentee after: ZHONGYI TECHNOLOGY Co.,Ltd. Address before: 250353 University Road, Changqing District, Ji'nan, Shandong Province, No. 3501 Patentee before: Qilu University of Technology (Shandong Academy of Sciences) |
|
| PE01 | Entry into force of the registration of the contract for pledge of patent right | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right |
Denomination of invention: A Signal Detection Method Based on Model Driven Deep Learning Granted publication date: 20220517 Pledgee: Weihai City Commercial Bank Limited by Share Ltd. high tech branch Pledgor: ZHONGYI TECHNOLOGY Co.,Ltd. Registration number: Y2024980002025 |