Method for implementing uplink and downlink channel reciprocity, communication node, and storage medium
A method for implementing uplink and downlink channel reciprocity, a communication node, and a storage medium are provided. The method for implementing uplink and downlink channel reciprocity includes: performing channel estimation in real time according to a received reference signal to obtain a channel estimation result; inputting the channel estimation result into a pre-trained neural network model, to output a channel adaptive matrix; and performing channel adaptation processing on a transmission signal by applying the channel adaptive matrix to the transmission signal.
1. A method for implementing uplink and downlink channel reciprocity, applied to a communication node, the method comprising:
performing channel estimation in real time according to a received reference signal to obtain a channel estimation result;
inputting the channel estimation result into a pre-trained neural network model, to output a channel adaptive matrix; and
performing channel adaptation processing for a transmission signal by applying the channel adaptive matrix to the transmission signal.
2. The method according to claim 1 , wherein, in a case where the communication node is a first communication node, the reference signal is an uplink reference signal; the channel estimation is an uplink channel estimation; the channel estimation result is an uplink channel estimation result; the channel adaptive matrix is a downlink channel adaptive matrix; and the channel adaptation processing is a downlink channel adaptation processing.
3. The method according to claim 1 , wherein, in a case where the communication node is a second communication node, the reference signal is a downlink reference signal; the channel estimation is a downlink channel estimation; the channel estimation result is a downlink channel estimation result; the channel adaptive matrix is an uplink channel adaptive matrix; and the channel adaptation processing is an uplink channel adaptation processing.
4. The method according to claim 1 , wherein the neural network model comprises: a complex neural network model; the complex neural network model comprises: a complex convolutional neural network model; and the complex convolutional neural network comprises: a complex residual neural network model.
5. The method according to claim 4 , wherein, in a case where the neural network model is the complex neural network model, the complex convolutional neural network model or a complex residual network model, a complex error back propagation algorithm is used to train the neural network model.
6. The method according to claim 5 , wherein the complex error back propagation algorithm comprises:
a computation expression of a partial derivative or sensitivity of a loss function E to a complex weight and W u,v j,k a complex bias b v k of a k-th full-connected layer in the complex neural network model shown as follows:
δ
W
u
,
v
j
,
k
=
Δ
∂
E
∂
W
u
,
v
j
,
k
=
conj
(
x
u
j
)
·
conj
(
f
′
(
x
u
j
)
·
δ
x
v
k
δ
b
v
k
=
Δ
∂
E
∂
b
v
k
=
δ
v
k
wherein, W u,v j,k represents a complex weight of a connection from a u-th neuron of a j-th layer to a v-th neuron of the k-th layer in the complex neural network model; b v k represents a bias of the v-th neuron of the k-th layer in the complex neural network model; x u j and x v k represent a complex output signal of the u-th neuron of the j-th layer and a complex output signal of the v-th neuron of the k-th layer in the complex neural network model respectively; f′(●) represents a derivative of a neuron activation function to an input signal; and conj(⋅) represents a conjugate operation on a complex number.
7. The method according to claim 5 , wherein the complex error back propagation algorithm comprises:
a computation expression of a partial derivative or sensitivity of a loss function E to a complex weight {right arrow over (Λ)} q k and a complex bias b q k of a q-th kernel of a k-th convolutional layer in the complex neural network model shown as follows:
δ
¯
Λ
→
q
k
=
Δ
∂
E
∂
Λ
→
q
k
=
∑
p
=
1
P
C
o
nv
(
Fliplr
(
conj
(
X
→
p
j
)
,
δ
¯
X
→
q
k
)
δ
b
q
k
=
Δ
∂
E
∂
b
q
k
=
∑
v
=
1
V
δ
x
q
,
v
k
where, {right arrow over (X)} p j , represents a p-th complex vector outputted by a previous layer, i.e., a j-th layer of the k-th convolutional layer; {right arrow over (X)} q k and represents a q-th complex vector outputted by the k-th convolutional layer; {right arrow over (δ)}x q k =[δ x q,l k , . . . , δ x q,v k , . . . , δ x q,V k ] represents a partial derivative or sensitivity of the loss function E to the above {right arrow over (X)} q k Conv(●) represents a convolutional operation, and Fliplr(●) represents flipping a transpose of an input vector, and p represents a total number of complex vectors outputted from the j-th convolutional layer.
8. The method according to claim 1 , wherein a training process for the neural network model comprises:
dividing wireless channel scenarios of a frequency division duplexing (FDD) system where the communication node is located into N types of scenarios, wherein N is a positive integer greater than or equal to 1;
performing uplink channel estimation and downlink channel estimation for each type of scenarios according to known training symbols to obtain uplink channel estimation results and first downlink channel estimation results; and
sequentially inputting an uplink channel estimation result of each entry of entries in a data set corresponding to the each type of scenarios into the neural network model to output a respective second downlink channel estimation result, until an error of the respective second downlink channel estimation result with respect to a first downlink channel estimation result served as a training tag meets a preset accuracy, wherein the entries are data pairs each composed of a uplink channel estimation result of the uplink channel estimation results and a respective first downlink channel estimation result of the first downlink channel estimation results.
9. The method according to claim 1 , wherein a training process for the neural network model comprises:
performing uplink channel estimation and downlink channel estimation in real time according to known training symbols, to obtain uplink channel estimation results and first downlink channel estimation results; and
sequentially inputting each uplink channel estimation result of uplink channel estimation results into the neural network model, and outputting a respective second downlink channel estimation result, until an error of the respective second downlink channel estimation result with respect to a first downlink channel estimation result served as a training tag meets a preset accuracy.
10. A communication node, comprising: a communication module, a memory, and at least one processor; wherein
the communication module is configured to perform communication interaction between at least two communication nodes;
the memory is configured to store at least one program; and
the at least one program, when being executed by the at least one processor, causes the at least one processor to implement:
performing channel estimation in real time according to a received reference signal to obtain a channel estimation result;
inputting the channel estimation result into a pre-trained neural network model, to output a channel adaptive matrix; and
performing channel adaptation processing for a transmission signal by applying the channel adaptive matrix to the transmission signal.
11. The communication node according to claim 10 , wherein, in a case where the communication node is a first communication node, the reference signal is an uplink reference signal; the channel estimation is an uplink channel estimation; the channel estimation result is an uplink channel estimation result; the channel adaptive matrix is a downlink channel adaptive matrix; and the channel adaptation processing is a downlink channel adaptation processing.
12. The communication node according to claim 10 , wherein, in a case where the communication node is a second communication node, the reference signal is a downlink reference signal; the channel estimation is a downlink channel estimation; the channel estimation result is a downlink channel estimation result; the channel adaptive matrix is an uplink channel adaptive matrix; and the channel adaptation processing is an uplink channel adaptation processing.
13. The communication node according to claim 10 , wherein the neural network model comprises: a complex neural network model; the complex neural network model comprises: a complex convolutional neural network model; and the complex convolutional neural network comprises: a complex residual neural network model.
14. The communication node according to claim 13 , wherein, in a case where the neural network model is the complex neural network model, the complex convolutional neural network model or a complex residual network model, a complex error back propagation algorithm is used to train the neural network model.
15. The communication node according to claim 14 , wherein the complex error back propagation algorithm comprises:
a computation expression of a partial derivative or sensitivity of a loss function E to a complex weight W u,v j,k and a complex bias b v k of a k-th full-connected layer in the complex neural network model shown as follows:
δ
W
u
,
v
j
,
k
=
Δ
∂
E
∂
W
u
,
v
j
,
k
=
conj
(
x
u
j
)
·
conj
(
f
′
(
x
u
j
)
·
δ
x
v
k
δ
b
v
k
=
Δ
∂
E
∂
b
v
k
=
δ
v
k
wherein, W u,v j,k represents a complex weight of a connection from a u-th neuron of a j-th layer to a v-th neuron of the k-th layer in the complex neural network model; b v k represents a bias of the v-th neuron of the k-th layer in the complex neural network model; x u j and x v k represent a complex output signal of the u-th neuron of the j-th layer and a complex output signal of the v-th neuron of the k-th layer in the complex neural network model respectively; f′(●) represents a derivative of a neuron activation function to an input signal; and conj(⋅) represents a conjugate operation on a complex number.
16. The communication node according to claim 14 , wherein the complex error back propagation algorithm comprises:
a computation expression of a partial derivative or sensitivity of a loss function E to a complex weight Λ q k and a complex bias b q k of a q-th kernel of a k-th convolutional layer in the complex neural network model shown as follows:
δ
¯
Λ
→
q
k
=
Δ
∂
E
∂
Λ
→
q
k
=
∑
p
=
1
P
C
o
nv
(
Fliplr
(
conj
(
X
→
p
j
)
,
δ
¯
X
→
q
k
)
δ
b
q
k
=
Δ
∂
E
∂
b
q
k
=
∑
v
=
1
V
δ
x
q
,
v
k
where, {right arrow over (X)} p j represents a p-th complex vector outputted by a previous layer, i.e., a j-th layer of the k-th convolutional layer; {right arrow over (X)} q k and represents a q-th complex vector outputted by the k-th convolutional layer; {right arrow over (δ)}x q k =[δ x q,l k , . . . , δ x q,v k , . . . , δ x q,V k ] represents a partial derivative or sensitivity of the loss function E to the above {right arrow over (X)} q k , Conv(●) represents a convolutional operation, and Fliplr(●) represents flipping a transpose of an input vector, and p represents a total number of complex vectors outputted from the j-th convolutional layer.
17. The communication node according to claim 10 , wherein a training process for the neural network model comprises:
dividing wireless channel scenarios of a frequency division duplexing (FDD) system where the communication node is located into N types of scenarios, wherein Nis a positive integer greater than or equal to 1;
performing uplink channel estimation and downlink channel estimation for each type of scenarios according to known training symbols to obtain uplink channel estimation results and first downlink channel estimation results; and
sequentially inputting an uplink channel estimation result of each entry of entries in a data set corresponding to the each type of scenarios into the neural network model to output a respective second downlink channel estimation result, until an error of the respective second downlink channel estimation result with respect to a first downlink channel estimation result served as a training tag meets a preset accuracy, wherein the entries are data pairs each composed of a uplink channel estimation result of the uplink channel estimation results and a respective first downlink channel estimation result of the first downlink channel estimation results.
18. The communication node according to claim 10 , wherein a training process for the neural network model comprises:
performing uplink channel estimation and downlink channel estimation in real time according to known training symbols, to obtain uplink channel estimation results and first downlink channel estimation results; and
sequentially inputting each uplink channel estimation result of uplink channel estimation results into the neural network model, and outputting a respective second downlink channel estimation result, until an error of the respective second downlink channel estimation result with respect to a first downlink channel estimation result served as a training tag meets a preset accuracy.
19. A non-transitory computer-readable storage medium, storing a computer program, wherein the computer program, when being executed by a processor, implements:
performing channel estimation in real time according to a received reference signal to obtain a channel estimation result;
inputting the channel estimation result into a pre-trained neural network model, to output a channel adaptive matrix; and
performing channel adaptation processing for a transmission signal by applying the channel adaptive matrix to the transmission signal.