Methods and systems for performing radio-frequency signal noise reduction in the absence of noise models

• US 10,291,268 B1
• Filed: 07/25/2017
• Issued: 05/14/2019
• Est. Priority Date: 07/25/2017
• Status: Active Grant

• Current Assignee: the united states of america as represented by the secretary of the navy
• Original Assignee: The United States of America As Represented By The Secretary of Agriculture
• Inventors: Migliori, Benjamin J., Gebhardt, Daniel J., Walton, Michael W., Straatemeier, Logan M.
• Primary Examiner(s): Trinh, Sonny
• Application Number:
  US15/659,298
• Time in Patent Office: 658 Days
• Field of Search: 455 631, 455 6713, 4551142, 4552772, 4552778, 706 21, 706 15, 706 31, 706 32
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: H04B 1/0475 with means for limiting noise, interference or distortion H04B1/0483 takes precedence
• CPC Class Codes: H04B 1/12 Neutralising, balancing, or compensation arrangements
• CPC Class Codes: H04B 1/70712 with demodulation by means of convolvers, e.g. of the SAW type SAW convolvers in general G06G7/195
• CPC Class Codes: H04B 1/71055 using minimum mean squared error [MMSE] detector
• CPC Class Codes: H04B 1/7176 Data mapping, e.g. modulation

Abstract

Time-varying input signals are denoised by a neural network. The neural network learns features associated with noise added to reference signals. The neural network recognizes features of noisy time-varying input signals mixed with the noise that at least partially match at least some of the features associated with the noise. The neural network predicts denoised time-varying output signals that correspond to the time-varying input signals based on the recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise.

Claims

1. A method for performing denoising of time-varying input signals, comprising:

a) learning, by a neural network, features associated with noise added to reference signals;

b) recognizing, by the neural network, features of noisy time-varying input signals mixed with the noise that at least partially match at least some of the features associated with the noise;

c) predicting, by the neural network, denoised time-varying output signals that correspond to the time-varying input signals based on the recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise.

2. The method of claim 1, wherein the reference signals include a series of known random signals received in a transmission environment including the noise, and the time-varying input signals include radio frequency signals received in the transmission environment including the noise.

3. The method of claim 1, wherein the neural network includes at least one layer, and step a) comprises unsupervised training including assigning, by the neural network, weights of filters of the layer of the neural network based on the features associated with the noise added to the reference signals.

4. The method of claim 3, wherein unsupervised training further comprises adjusting the assigned weights of the filters of the layer of the neural network.

5. The method of claim 4, wherein adjusting the assigned weights of the filters includes predicting, by the neural network, denoised output reference signals that correspond to the reference signals without noise, calculating a loss function corresponding to an error between the reference signals and the denoised output reference signals predicted by the neural network and adjusting the weights of the filters such that the loss function is minimized.

6. The method of claim 5, wherein the assigned weights are repeatedly adjusted to minimize the loss function.

7. The method of claim 3, wherein step b) comprises convolving the noisy time-varying input signals using the filters with the assigned weights to produce feature maps representing the recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise and wherein step c) comprises deconvolving the feature maps using transposed versions of the filters with the assigned weights to predict the denoised time-varying output signals that correspond to the time-varying input signals.

8. The method of claim 7, further comprising pooling the feature maps to produce pooled maps indicating maximum values within pooling regions of the feature maps and switch variables indicating the locations of the maximum values with the pooling regions of the feature maps and unpooling the pooled maps using the switch variables to produce the feature maps.

9. A system for automatically denoising time-varying input signals, comprising:

at least one autoencoder including at least one convolutional layer with filters having assigned weights and at least one deconvolutional layer with transposed versions of the filters having the assigned weights, wherein the autoencoder is configured to;

during a training phase;

learn features associated with noise added to reference signals;

and

during a denoising phase;

recognize features of noisy time-varying input signals mixed with the noise that at least partially match at least some of the features associated with the noise; and

predict denoised time-varying output signals that correspond to the time-varying input signals based on the recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise.

10. The system of claim 9, wherein the convolutional layer convolves the noisy time-varying input signals using the filters with the assigned weights to produce feature maps representing the recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise, and the deconvolutional layer deconvolves the feature maps using the transposed versions of the filters with the assigned weights to predict the denoised time-varying output signals that correspond to the time-varying input signals.

11. The system of claim 9, further comprising multiple stacked autoencoders including multiple convolutional layers and multiple deconvolutional layers.

12. The system of claim 9, wherein, during the training phase, the autoencoder learns the features associated with the noise by predicting denoised output reference signals that correspond to the reference signals without noise, calculating a loss function corresponding to an error between the reference signals and the denoised output reference signals predicted by the neural network, and adjusting the assigned weights of the filters such that the loss function is minimized.

13. The system of claim 12, wherein the loss function is a mean squared error function.

14. A device for denoising time-varying input signals, comprising:

a processor; and

a memory having stored thereon instructions which, when executed by the processor, cause the processor to perform steps comprising;

a) performing unsupervised training of a neural network having at least one layer, including assigning weights of filters of the layer of the neural network based on the features associated with the noise added to the reference signals;

b) convolving the noisy time-varying input signals using the filters with the assigned weights to produce feature maps representing recognized features of the noisy time-varying input signals that at least partially match at least some of the features associated with the noise; and

c) deconvolving the feature maps using transposed versions of the filters with the assigned weights to predict denoised time-varying output signals that correspond to the time-varying input signals.

15. The device of claim 14, wherein the reference signals include a series of known random signals received in a transmission environment including the noise, and the time-varying input signals include radio frequency signals received in the transmission environment including the noise.

16. The device of claim 14, wherein unsupervised training further comprises adjusting the assigned weights of the filters of the layer of the neural network.

17. The device of claim 14, wherein adjusting the assigned weights of the filters includes predicting denoised output reference signals that correspond to the reference signals without noise, calculating a loss function corresponding to an error between the reference signals and the predicted denoised output reference signals and adjusting the weights of the filters such that the loss function is minimized.

18. The device of claim 17, wherein the assigned weights of the filters are repeatedly adjusted to minimize the loss function.

19. The device of claim 18, wherein the loss function is a mean squared error function.

20. The device of claim 14, wherein the instructions, when executed by the processor, further cause the processor to perform:

pooling the feature maps to produce pooled maps indicating maximum values within pooling regions of the feature maps switch variables indicating the locations of the maximum values with the pooling regions of the feature maps; and

unpooling the pooled maps using the switch variables to produce the feature maps.