Granted Patent B2

US 12,273,221 · App. 18/395,704 · Granted Apr 8, 2025

Inventors: Xiaohua Li (Johnson City, NY); Robert Thompson (Quakertown, PA)

H04L25/03165G06N3/04G06N3/08H03F1/32H03F3/20H04B1/16H03F2200/451

Abstract

Claims (245)

1. A method of compensating for non-linear distortion from a power amplifier having a non-linearity, comprising:

digitizing a non-linearly distorted signal from a radio frequency receiver;

transforming the digitized non-linearly distorted signal into a series expansion comprising at least third order terms with memory with at least one automated processor, each respective term representing an incremental non-linearity order and an associated delay;

processing the series expansion with a deep neural network, comprising at least one dropout layer, trained with respect to the non-linear distortion to produce an equalized output signal; and

extracting a set of information symbols from the equalized output signal with a demodulator.

2. The method according to claim 1 , wherein the neural network is trained with training data comprising a set of small amplitude training signals to estimate a channel response independent of the power amplifier and a set of large amplitude training signals associated with the power amplifier non-linearity.

3. The method according to claim 1 , wherein the non-linearly distorted signal comprises an orthogonal frequency division multiplexed signal.

4. The method according to claim 1 , wherein the series expansion of at least third order with memory comprises a Volterra series expansion.

5. The method according to claim 4 ,

wherein the terms of the Volterra series expansion are defined by:

y

⁡

(

n

)

=

∑

d

=

0

D

∑

k

=

0

P

b

k

⁢

d

⁢

x

⁡

(

n

-

d

)

⁢

❘

"\[LeftBracketingBar]"

x

⁡

(

n

-

d

)

❘

"\[RightBracketingBar]"

k

-

1

;

and

the neural network is a multi-layer feedforward neural network trained with respect to the channel non-linear distortion to produce:

z

⁡

(

n

)

=

∑

k

=

1

P

∑

d

1

=

0

D

…

⁢

∑

d

k

=

0

D

f

d

1

,

…

,

d

k

⁢

∏

i

=

1

k

r

⁡

(

n

-

d

i

)

,

where:

y(n) represents the non-linearly distorted communication;

z(n) is the equalized output signal;

d is a memory depth parameter;

D is a total memory length;

k is a respective nonlinearity order;

P is a total nonlinearity order;

k is a nonlinear order;

b kd are nonlinear response parameters; and

r(n) is a channel response to the non-linearly distorted signal y(n).

6. The method according to claim 1 , wherein the series expansion of at least third order with memory comprises at least fifth order terms, and the neural network has at least two convolutional network layers.

7. The method according to claim 1 , wherein the neural network comprises at least three hidden layers, each hidden layer comprising at least 10 feature maps, and a fully connected layer subsequent to the at least three hidden layers.

8. The method according to claim 1 , wherein the non-linearly distorted signal comprises a frequency division multiplexed radio frequency modulated set of symbols non-linearly distorted by the power amplifier, and communicated through a non-linear channel.

9. The method according to claim 1 , further comprising adaptively updating the neural network to reduce an error rate of the demodulated set of symbols.

10. The method according to claim 1 , further comprising producing the equalized output signal as an at least two dimensional tensor over time, wherein the neural network comprises at least three one-dimensional convolutional layers.

11. The method according to claim 10 , wherein the at least three one-dimensional convolutional layers each use a sigmoid activation function.

12. The method according to claim 10 , wherein the at least three one-dimensional convolutional layers of the neural network are followed by a first fully-connected layer, the at least one dropout layer, and a fully-connected output layer.

13. The method according to claim 12 , wherein the first fully connected layer uses a sigmoid activation function, and the fully connected output layer uses a linear activation function.

14. A system for compensating for non-linear distortion, comprising:

a radio frequency receiver configured to receive a non-linearly distorted signal;

a memory configured to store a digitized representation of the non-linearly distorted signal;

at least one automated processor configured to transform the digitized representation of the non-linearly distorted signal into a series expansion comprising at least third order terms with memory, each respective term representing an incremental non-linearity order and an associated delay;

an artificial neural network comprising at least one dropout layer, configured to receive the series expansion, and trained with respect to the non-linear distortion to produce a reduced distortion output;

an output configured to communicate a representation of the reduced distortion output; and

a demodulator configured to extract information symbols from the reduced distortion output.

15. The system according to claim 14 , wherein the neural network is trained with training data comprising a set of small amplitude training signals to estimate a channel response independent of the power amplifier and a set of large amplitude training signals associated with the non-linear distortion.

16. The system according to claim 14 , wherein the radio frequency receiver comprises a software-defined radio having a wideband receiver bank, a digital physical/media access control (PHY/MAC) layer receiver, and a protocol analyzer.

17. The system according to claim 14 , wherein the series expansion of at least third order with memory comprises a Volterra series expansion comprising at least fifth order terms, and the terms of the Volterra series expansion are defined by:

y

⁡

(

n

)

=

∑

d

=

0

D

∑

k

=

0

P

b

k

⁢

d

⁢

x

⁡

(

n

-

d

)

⁢

❘

"\[LeftBracketingBar]"

x

⁡

(

n

-

d

)

❘

"\[RightBracketingBar]"

k

-

1

;

and

the neural network is a multi-layer feedforward neural network trained with respect to the channel non-linear distortion to produce:

z

⁡

(

n

)

=

∑

k

=

1

P

∑

d

1

=

0

D

…

⁢

∑

d

k

=

0

D

f

d

1

,

…

,

d

k

⁢

∏

i

=

1

k

⁢

r

⁡

(

n

-

d

i

)

,

where:

y(n) represents the non-linearly distorted signal;

z(n) is the reduced distortion output;

d is a memory depth parameter;

D is a total memory length;

k is a respective nonlinearity order;

P is a total nonlinearity order;

k is a nonlinear order;

b kd are nonlinear response parameters; and

r(n) is a channel response to the non-linearly distorted signal y(n).

18. The system according to claim 14 , wherein the neural network comprises at least three hidden layers, each hidden layer comprising at least 10 feature maps, and a fully connected layer subsequent to the at least three hidden layers, and the at least one process to is further configured to adaptively update the neural network to reduce an error of the reduced distortion output.

19. The system according to claim 14 , wherein the reduced distortion output comprises an at least two dimensional tensor over time, and the neural network comprises at least three one-dimensional convolutional layers each use a sigmoid activation function, followed by a first fully-connected layer which uses a sigmoid activation function, the at least one dropout layer, and a fully-connected output layer which uses a linear activation function.

20. A processor configured to compensate for distortion from a non-linear process, comprising:

a radio frequency receiver configured to receive a distorted signal communicating a series of symbols;

at least one automated processor configured to transform the distorted signal into a truncated Volterra series expansion comprising a series of terms of at least third order and comprising a sum of multidimensional convolutions of at least third order, each with an associated time delay component;

a memory configured to store the Volterra series expansion;

a multi-layer feedforward deep neural network comprising a plurality of convolutional neural network layers and at least one dropout layer, trained with respect to a distortion of the distorted signal, and adaptively updated dependent on output error, configured to process the Volterra series expansion, and to produce an output having reduced distortion with respect to the distorted signal; and

a demodulator configured to extract the series of symbols from the output having reduced distortion.

Continuity (5)

References Cited (400)

Cited By (1)

US 12,456,994