Deep learning based multi-branch joint detection method for cooperative communication

Abstract

The present invention relates to the technical field of signal processing for cooperative communication, and more specifically, to a multi-branch joint detection method based on deep learning for cooperative communication, which can effectively process multi-node signals and eliminate inter-code interference in the communication process. The method is characterized in that a multi-branch joint detector based on deep learning is established to directly process signals from source nodes and relay nodes, wherein in the multi-branch joint detector based on deep learning, a DNN is set to SD The input layer (layer 1) of the neural network is composed of c SD (1) It means that DNN RD The input of the neural network input layer is represented by c RD (1) Different from the existing single-branch method based on deep learning, the multi-branch joint detector based on deep learning proposed in the present invention can jointly realize the merging of multiple branches, and the adjustment of each branch tap is based on the total output of each branch, and then based on the calculated total error, which can effectively eliminate inter-code interference.

Description

Technical field:

[0001] The present invention relates to the technical field of signal processing for collaborative communication, and specifically to a multi-branch joint detection method based on deep learning for collaborative communication, which can effectively process multi-node signals and eliminate inter-code interference during the communication process. Background technology:

[0002] Collaborative communication has many advantages, including improving the link reliability of the system. However, the improvement of link reliability depends to a certain extent on whether the signals from multiple nodes can be effectively processed. In particular, in some environments, the channels are frequency selective, and the effective elimination of inter-symbol interference (ISI) must also be considered.

[0003] At present, most studies consider the merging and equalization of multi-node signals independently, and the equalization usually adopts the traditional equalization algorithm. In order to achieve better collaborative performance, a multi-branch joint detector (MJD, Multi-BranchJoint Detector) has been proposed in the prior art. The detector jointly realizes multi-branch merging and equalization, which can achieve significant performance improvement compared with the separate multi-branch merging and equalization. However, its equalization adopts the traditional equalization algorithm and linear structure, and its ability to eliminate inter-code interference is limited. In recent years, artificial intelligence algorithms have made it possible to solve this problem, especially the introduction of deep learning algorithms, because they have excellent nonlinear fitting capabilities. In the prior art, multi-layer perceptrons (MLP, Multilayer Perceptrons) were first used to equalize nonlinear channels, but its structure is relatively simple and can only handle the situation where there are very few multipath channels. In addition, some researchers have used deep neural network (DNN), long short-term memory network (LSTM), and convolutional neural network (CNN) to study channel equalization in point-to-point orthogonal frequency division multiplex (OFDM) systems, but these methods are not ideal under single-carrier conditions.

[0004] As far as we know, there is currently no research on the joint implementation of multi-branch merging and ISI elimination based on deep learning. Summary of the invention:

[0005] In view of the shortcomings and deficiencies of the prior art, the present invention proposes a multi-branch joint detection method based on deep learning for collaborative communication, which can jointly realize the merging of multiple branches and the elimination of inter-code interference, and has nonlinear capabilities due to the combination of deep learning algorithms, and can achieve better system performance.

[0006] The present invention is achieved by the following measures:

[0007] A multi-branch joint detection method based on deep learning for cooperative communication is applied to a cooperative communication system including a source node S, a destination node D and a relay node R. The system relay works in a decoding and forwarding mode. The cooperation includes a broadcast phase and a relay phase. In the broadcast phase, the signal r received by the relay node R SR (n), the signal r received by the destination node D SD (n) is expressed as follows:

[0008] r SR (n) = s(n) * h SR (n)+n SR (n) (1)

[0009] r SD (n) = s(n) * h SD (n)+n SD (n) (2)

[0010] Where s(n) is an equally probable uncorrelated binary phase shift keying (BPSK) signal, * represents convolution, and n SR (n) and n SD (n) has a mean of zero and a variance of N 0,R 、N 0,D Additive Gaussian white noise, h SR (n), h SD (n) represents the impulse response of the channel between the source node S and the relay node R, and the source node S and the destination node D respectively; the channels between all nodes are assumed to be frequency selective channels, and the channel impulse response is modeled by the following formula:

[0011]

[0012] where j∈{SR,SD,RD}, L j represents the length of the channel impulse response, σ describes the channel delay spread time, P R Indicates channel h j (n) The average power of the multipath components;

[0013] In the relay stage, assuming that the relay node can perfectly restore the information sent by the source node, the signal received by the destination node D from the relay node R is expressed as:

[0014] r RD (n) = s(n) * h RD (n)+n RD (n) (4)

[0015] where h RD (n) represents the impulse response of the channel between the relay node R and the destination node D, n RD (n) has a mean of zero and a variance of N R,D Additive Gaussian white noise;

[0016] It is characterized in that a multi-branch joint detector based on deep learning is established to directly process signals from source nodes and relay nodes, wherein in the multi-branch joint detector based on deep learning, DNN SD The input layer (layer 1) of the neural network is composed of c SD (1) It means that DNN RD The input of the neural network input layer is represented by c RD (1) , then from r SD (n) and r RD (n) The obtained c SD (1) and c RD (1) The expression is represented by the following formula:

[0017] c SD (1) =[r SD (n),...,r SD (nM SD +1)] (5)

[0018] c RD (1) =[r RD (n),...,r RD (nM RD +1)] (6)

[0019] M SD and M RD Represents DNN SD and DNN RD The size of the number of neurons in the input layer;

[0020] Assume that the number of network layers of the DNNs of the two branches is L, and set DNN SD The lth layer (l=1,...,L-2) has N l neurons, and the output vector of the lth layer is c SD (l) (The length of this vector is Nl ), then the output of the l+1th layer is

[0021] c SD (l+1) =f (l) (c SD (l) )=f(θ (l) c SD (l) +v (l) ) (7)

[0022] Among them, θ (l) For DNN SD The weight matrix of the lth layer, and (i=1,...,N l+1 ,j=1,...,N l ), v (l) For DNN SD The bias vector of the lth layer, and (i=1,...,N l ), f(·) is the activation function, which uses the hyperbolic tangent (tanh) function and is defined as follows:

[0023]

[0024] Assume DNN RD The lth layer (l=1,...,L-2) of l neurons, and the output vector of the lth layer is c RD (l) (The length of this vector is N l ), then the output of the l+1th layer is

[0025] c RD (l+1) =f (l) (c RD (l) )=f(λ (l) c RD (l) +k (l) ) (9)

[0026] Among them, λ (l) For DNN RD The weight matrix of the lth layer, and (i=1,...,N l+1 ,j=1,...,N l ), k (l) For DNN RD The bias vector of the lth layer, and (i=1,...,N l );

[0027] DNN SD and DNN RD The output of the L-1 layer (the layer before the output layer) is expressed as:

[0028] c SD (L-1) =f (L-1) (...f (1) (c SD (1) )) (10)

[0029] c RD (L-1) =f (L-1) (...f (1) (c RD (1) )) (11).

[0030] In order to reduce the amount of calculation, the present invention needs to reduce the scale of the neural network. Therefore, in each training, several groups are randomly selected from all training samples, and then a training is completed. Different from the single-branch structure, the network outputs of the two branches are first merged and then processed by the activation function f(·) to obtain the total output of the detector DL-MJD:

[0031]

[0032] in is a vector of length N, N represents the number of samples for each training.

[0033] In the present invention, the cost function of the multi-branch joint detector DL-MJD based on deep learning can be constructed from the estimated total output as follows:

[0034]

[0035] where p(j) is the corresponding estimated value The known training data, n = 1, ..., N, N is the number of training data, which is the same as the number of samples in each training.

[0036] The smaller the value of the cost function in the present invention, the smaller the estimated value The closer it is to the expected value s(n), therefore, the cost function is minimized to optimize the neural network:

[0037] L(θ,v,λ,k) opt =minL(θ,v,λ,k) (14)

[0038] In order to obtain an adaptive algorithm that minimizes the cost function, the present invention completes the adaptive optimization of the neural network based on the gradient descent method. Different from the single-branch method, here the multi-branch signal is processed, and the total error is calculated using the total output of all branches to optimize and update the neural network weight vector of each branch so that the cost function (13) tends to the optimal value. According to the gradient descent method, the update of the network parameters of each branch can be obtained by the following formula:

[0039]

[0040]

[0041]

[0042]

[0043] Among them, α represents the learning rate,

[0044] For the derivative in formula (15), since the error is back-propagated, the partial derivative of the L-1th layer is first derived, and the result is shown in the following formula:

[0045]

[0046]

[0047] Where ⊙ represents the Hadamard product, (·) T Represents matrix transpose.

[0048] In order to obtain the partial derivatives of other values ​​of l in (15) (l = 1, ..., L-2), DNN SD The error term δ of the L-1th layer neurons L-1 Defined as:

[0049]

[0050] Similarly, after derivation, DNN SD The error term δ of the lth layer of neurons l (l=1,...,L-2) is obtained from the following formula:

[0051]

[0052] From formula (19), we can get:

[0053]

[0054]

[0055] also, and It is expressed in the following format:

[0056]

[0057]

[0058]

[0059] in, For DNN RD Substituting (20), (21), (23) and (24) into (15), the adaptive algorithm of DL-MJD is obtained as follows:

[0060] θ (l) =θ (l) -αδ l (c SD l-1 ) T (l=1,...L-1)

[0061] v (l) =v (l) -αδ l (l=1,...L-1)

[0062]

[0063]

[0064] Different from the existing single-branch method based on deep learning, the multi-branch joint detector based on deep learning proposed in the present invention can jointly realize the merging of multiple branches, and the adjustment of each branch tap is based on the total output of each branch, and then based on the calculated total error, which can effectively eliminate inter-code interference. Description of the drawings:

[0065] Attached Figure 1 It is a structural diagram of the multi-branch joint detector DL-MJD based on deep learning in the present invention.

[0066] Attached Figure 2 It is the bit error rate performance of DL-MJD at different numbers of layers in Example 1, where SNR = 12dB.

[0067] Attached Figure 3 It is a bit error rate performance comparison curve chart of different algorithms in Example 1. Specific implementation method:

[0068] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0069] The present invention considers a cooperative communication system including a source node S, a destination node D and a relay node R. The study assumes that the relay works in a decode-and-forward mode (DF).

[0070] There are two stages to complete the collaboration. The first is the broadcast stage, that is, the source node broadcasts the signal to the relay node and the destination node respectively; the second is the relay stage, that is, the relay node forwards the received signal to the destination node by amplifying and forwarding.

[0071] In the broadcast phase, the signal r received by the relay node R SR (n), the signal r received by the destination node D SD (n) can be expressed as follows:

[0072] r SR (n) = s(n) * h SR (n)+n SR (n) (1)

[0073] r SD (n) = s(n) * h SD (n)+n SD (n) (2)

[0074] Where s(n) is an equally probable uncorrelated binary phase shift keying (BPSK) signal, * represents convolution, and n SR (n) and n SD (n) has a mean of zero and a variance of N 0,R 、N 0,D Additive Gaussian white noise, h SR (n), h SD (n) respectively represent the impulse responses of the channels between the source node S and the relay node R, and between the source node S and the destination node D. In the present invention, the channels between all nodes are assumed to be frequency selective channels, and the channel impulse responses are modeled by the following formula:

[0075]

[0076] where j∈{SR,SD,RD}, L j represents the length of the channel impulse response, σ describes the channel delay spread time, P R Indicates channel h j (n) The average power of the multipath components.

[0077] In the relay stage, after the relay node receives the signal from the source node, it first decodes the received signal, and then performs BPSK modulation on the decoded information like the source node, and then sends it to the destination node. Assuming that the relay node can perfectly restore the information sent by the source node, the signal received by the destination node D from the relay node R can be expressed as:

[0078] r RD (n) = s(n) * h RD (n)+n RD (n) (4)

[0079] where h RD (n) represents the impulse response of the channel between the relay node R and the destination node D, n RD (n) has a mean of zero and a variance of N R,D Additive white Gaussian noise.

[0080] In view of the nonlinear approximation capability of deep learning, the present invention proposes a deep learning-based multi-branch joint detector (DL-MJD, Deep Learning-based Multi-Branch Joint Detector). Different from the existing single-branch method based on deep learning, the detector can jointly realize the merging of multiple branches, and the adjustment of each branch tap is based on the total output of each branch, and then based on the calculated total error.

[0081] The structure of the proposed DL-MJD is as follows: Figure 1 As shown in Figure 2, the detector will directly process the signals from the source node and the relay node. SD The input layer (layer 1) of the neural network is composed of c SD (1) It means that DNN RD The input of the neural network input layer is represented by c RD (1) , then from r SD (n) and r RD (n) The obtained c SD (1) and c RD (1) The expression can be expressed as follows:

[0082] c SD (1) =[r SD (n),...,r SD (nM SD +1)] (5)

[0083] c RD (1) =[r RD(n),...,r RD (nM RD +1)] (6)

[0084] Among them, M SD and M RD Represents DNN SD and DNN RD The number of neurons in the input layer. Assume that the number of network layers of the two branches of DNN is L, and set DNN SD The lth layer (l=1,...,L-2) has N l neurons, and the output vector of the lth layer is c SD (l) (The length of this vector is N l ), then the output of the l+1th layer is

[0085] c SD (l+1) =f (l) (c SD (l) )=f(θ (l) c SD (l) +v (l) ) (7)

[0086] Among them, θ (l) For DNN SD The weight matrix of the lth layer, and (i=1,...,N l+1 ,j=1,...,N l ), v (l) For DNN SD The bias vector of the lth layer, and (i=1,...,N l ), f(·) is the activation function, where the hyperbolic tangent (tanh) function is used, which is defined as follows:

[0087]

[0088] Similarly, let DNN RD The lth layer (l=1,...,L-2) of l neurons, and the output vector of the lth layer is c RD (l) (The length of this vector is N l ), then the output of the l+1th layer is

[0089] c RD (l+1) =f (l) (c RD (l) )=f(λ(l) c RD (l) +k (l) ) (9)

[0090] Among them, λ (l) For DNN RD The weight matrix of the lth layer, and (i=1,...,N l+1 ,j=1,...,N l ), k (l) For DNN RD The bias vector of the lth layer, and (i=1,...,N l ).

[0091] Therefore, DNN SD and DNN RD The output of the L-1 layer (the layer before the output layer) can be expressed as:

[0092] c SD (L-1) =f (L-1) (...f (1) (c SD (1) )) (10)

[0093] c RD (L-1) =f (L-1) (...f (1) (c RD (1) )) (11)

[0094] In order to reduce the amount of calculation, the size of the neural network needs to be reduced. Therefore, in each training, several groups are usually randomly selected from all training samples and then sent to DL-MJD to complete a training. Different from the single-branch structure, the network outputs of the two branches are first merged and then processed by the activation function f(·) to obtain the total output of the detector DL-MJD:

[0095]

[0096] in is a vector of length N, N represents the number of samples for each training.

[0097] Based on the estimated total output, the cost function of DL-MJD can be constructed as follows:

[0098]

[0099] where p(j) is the corresponding estimated value The known training data, n = 1, ..., N, N is the number of training data, which is the same as the number of samples in each training.

[0100] Obviously, the smaller the cost function value, the smaller the estimated value The closer it is to the expected value s(n). Therefore, the cost function can be minimized to optimize the neural network:

[0101] L(θ,v,λ,k) opt =minL(θ,v,λ,k) (14)

[0102] In order to obtain an adaptive algorithm that minimizes the cost function, the present invention completes the adaptive optimization of the neural network based on the gradient descent method. Different from the existing single-branch method, here the multi-branch signal is processed, and the total error is calculated using the total output of all branches to optimize and update the neural network weight vector of each branch so that the cost function (13) tends to the optimal value. According to the gradient descent method, the update of the network parameters of each branch can be obtained by the following formula:

[0103]

[0104]

[0105]

[0106]

[0107] Among them, α represents the learning rate.

[0108] For the derivative in formula (15), since the error is back-propagated, the partial derivative of the L-1th layer is first derived, and the result is shown in the following formula:

[0109]

[0110]

[0111] Where ⊙ represents the Hadamard product, (·) T Represents matrix transpose.

[0112] In order to obtain the partial derivatives of other values ​​of l in (15) (l = 1, ..., L-2), we use DNN SD The error term δ of the L-1th layer neurons L-1 Defined as:

[0113]

[0114] Similarly, after derivation, DNN SD The error term δ of the lth layer of neuronsl (l=1,...,L-2) can be obtained from the following formula:

[0115]

[0116] From formula (19), we can get:

[0117]

[0118]

[0119] also, and The calculation is similar to the above and can be expressed in the following format:

[0120]

[0121]

[0122]

[0123] in, Substituting (20), (21), (23) and (24) into (15), we can get the adaptive algorithm of DL-MJD as follows:

[0124] θ (l) =θ (l) -αδ l (c SD l-1 ) T (l=1,...L-1)

[0125] v (l) =v (l) -αδ l (l=1,...L-1)

[0126]

[0127]

[0128] This section will illustrate the performance of the proposed DL-MJD scheme through simulation. For comparison, the simulation results of the existing single-branch detector (SBD, Single-Branch Detector), multi-branch joint detector (MJD, Multi-Branch Joint Detector) and deep learning-based single-branch detector (DL-SBD, Deep Learning-based Single-Branch Detector) methods are also given in the simulation.

[0129] The channel model used in the simulation is the model shown in equation (3), which is used to generate the channel impulse response between nodes. SD and DNN RD For example, the number of layers used in the simulation is 5, and the number of neurons N from the 1st to the 5th layer is l are set to 16, 16, 24, 36 and 1 respectively. In the simulation, the learning rate α of the algorithm is set to 0.005; the number of samples N selected for each training is 128, and the number of sampling points for each sample (M SD and M RD ) are both 16, the number of iterations is 1000; the length of the data packet sent is 2000 symbols, of which the first 500 are used as training sequences.

[0130] Figure 2 The bit error rate performance of the DL-MJD proposed in this invention is given under different numbers of layers. It can be seen from the figure that the change in the number of layers will affect the bit error rate performance of DL-MJD. As the number of layers increases, the bit error rate tends to be optimal, but when the number of layers increases further, the bit error rate tends to deteriorate. This is because the increase in the number of layers will cause the neural network to have an overfitting effect. At this time, increasing the number of network layers will not improve the fitting effect of the network, but will bring adverse effects. Figure 2 It can be seen that DL-MJD can achieve the best bit error rate performance when the number of layers is 5. In the following bit error rate performance comparison, the number of layers of DL-MJD will also be set to 5.

[0131] Figure 3 The bit error rate performance comparison between the proposed DL-MJD and the existing algorithms is given. Compared with SBD and DL-SBD, MJD and the proposed DL-MJD can achieve better bit error rate performance. This is because for MJD and DL-MJD, the addition of relay nodes enables the destination node to receive not only the signal from the source node, but also the signal forwarded by the relay node, and the receiving node can obtain a certain diversity gain after processing. In addition, it can be seen that DL-MJD can achieve better bit error rate performance than MJD. This is because the DL-MJD algorithm is implemented based on a deep learning algorithm. Compared with the linear structure of MJD, deep learning enables DL-MJD to have nonlinear fitting capabilities, which can more effectively eliminate the inter-code interference effect.

[0132] In summary, the present invention proposes a deep learning-based multi-branch joint detector (DL-MJD). The detector can jointly realize the merging of multi-branch signals and the elimination of inter-symbol interference in each branch. Moreover, the proposed method is based on the deep learning algorithm, so it can achieve better performance. The simulation results show that the proposed DL-MJD can achieve better performance than the existing algorithms.

Claims

1. A multi-branch joint detection method based on deep learning for cooperative communication is applied to a cooperative communication system including a source node S, a destination node D and a relay node R. The system relay works in a decoding and forwarding mode. The cooperation includes a broadcast phase and a relay phase. In the broadcast phase, the signal r received by the relay node R SR (n), the signal r received by the destination node D SD (n) is expressed as follows: r SR (n)=s(n)*h SR (n)+n SR (n) (1) r SD (n)=s(n)*h SD (n)+n SD (n) (2) Where s(n) is an equally probable uncorrelated binary phase shift keying (BPSK) signal, * represents convolution, and n SR (n) and n SD (n) has a mean of zero and a variance of N 0,R 、N 0,D Additive Gaussian white noise, h SR (n), h SD (n) represents the impulse response of the channel between the source node S and the relay node R, and the source node S and the destination node D respectively; the channels between all nodes are assumed to be frequency selective channels, and the channel impulse response is modeled by the following formula: where j∈{SR,SD,RD}, L j represents the length of the channel impulse response, σ describes the channel delay spread time, P R Indicates channel h j (n) the average power of the multipath components; In the relay stage, assuming that the relay node can perfectly restore the information sent by the source node, the signal received by the destination node D from the relay node R is expressed as: r RD (n)=s(n)*h RD (n)+n RD (n) (4) where h RD (n) represents the impulse response of the channel between the relay node R and the destination node D, n RD (n) has a mean of zero and a variance of N R,D Additive Gaussian white noise; It is characterized in that A multi-branch joint detector based on deep learning is established to directly process the signals from the source node and the relay node, wherein the multi-branch joint detector based on deep learning is set to SD The input of the input layer (layer 1) of the neural network is composed of c SD (1) It means that DNN RD The input of the neural network input layer is represented by c RD (1) , then from r SD (n) and r RD (n) The obtained c SD (1) and c RD (1) The expression is represented by the following formula: c SD (1) =[r SD (n),...,r SD (n-M SD +1)] (5) c RD (1) =[r RD (n),...,r RD (n-M RD +1)] (6) M SD and M RD Represents DNN SD and DNN RD The size of the number of neurons in the input layer; Assume that the number of network layers of the DNNs of the two branches is L, and set DNN SD The lth layer (l=1,...,L-2) has N l neurons, and the output vector of the lth layer is c SD (l) (The length of this vector is N l ), then the output of the l+1th layer is c SD (l+1) =f (l) (c SD (l) )=f(θ (l) c SD (l) +v (l) ) (7) Among them, θ (l) For DNN SD The weight matrix of the lth layer, and v (l) For DNN SD The bias vector of the lth layer, and As the activation function, the hyperbolic tangent (tanh) function is used, which is defined as follows: Assume DNN RD The lth layer (l=1,...,L-2) of l neurons, and the output vector of the lth layer is c RD (l) (The length of this vector is N l ), then the output of the l+1th layer is c RD (l+1) =f (l) (c RD (l) )=f(λ (l) c RD (l) +k (l) ) (9) Among them, λ (l) For DNN RD The weight matrix of the lth layer, and k (l) For DNN RD The bias vector of the lth layer, and DNN SD and DNN RD The output of the L-1 layer (the layer before the output layer) is expressed as: c SD (L-1) =f (L-1) (...f (1) (c SD (1) )) (10) c RD (L-1) =f (L-1) (...f (1) (c RD (1) )) (11)。 2. According to claim 1, a multi-branch joint detection method based on deep learning for collaborative communication is characterized in that: In order to reduce the amount of calculation, the size of the neural network needs to be reduced. Therefore, in each training, several groups are randomly selected from all training samples, and then a training is completed. Different from the single-branch structure, the network outputs of the two branches are first merged and then processed by the activation function f(·) to obtain the total output of the detector DL-MJD: in is a vector of length N, N represents the number of samples for each training.

3. The multi-branch joint detection method based on deep learning for collaborative communication according to claim 2, characterized in that: From the estimated total output, the cost function of the multi-branch joint detector DL-MJD based on deep learning can be constructed as follows: where p(j) is the corresponding estimated value The known training data, n = 1, ..., N, N is the number of training data, which is the same as the number of samples in each training.

4. The multi-branch joint detection method based on deep learning for collaborative communication according to claim 3, characterized in that: The smaller the cost function value, the smaller the estimated value. The closer it is to the expected value s(n), therefore, the cost function is minimized to optimize the neural network: L(θ,v,λ,k) opt =min L(θ,v,λ,k) (14) In order to obtain an adaptive algorithm that minimizes the cost function, the adaptive optimization of the neural network is completed based on the gradient descent method. According to the gradient descent method, the update of the network parameters of each branch can be obtained by the following formula: Among them, α represents the learning rate, For the derivative in formula (15), since the error is back-propagated, the partial derivative of the L-1th layer is first derived, and the result is shown in the following formula: Where ⊙ represents the Hadamard product, (·) T Represents matrix transpose; In order to obtain the partial derivatives of other values ​​of l in (15) (l = 1, ..., L-2), DNN SD The error term δ of the L-1th layer neurons L-1 Defined as: DNN SD The error term δ of the lth layer of neurons l (l=1,...,L-2) is obtained from the following formula: From formula (19), we can get: also, and It is expressed in the following format: in, For DNN RD Substituting (20), (21), (23) and (24) into (15), the adaptive algorithm of DL-MJD is obtained as follows: i (l) =θ (l) -ad l (c SD l-1 ) T (l=1,...L-1) v (l) =v (l) -αδ l (l=1,...L-1)

Images