Systems and methods for training multipath filtering systems

Patent metadata

Field	Value
Publication number	US-2016132768-A1
Application number	US-201414536798-A
Country	US
Kind code	A1
Filing date	Nov 10, 2014
Priority date	Nov 10, 2014
Publication date	May 12, 2016
Grant date	—

Abstract

A method for training a neural network to be configured to filter a multipath corrupted signal is provided. The method includes receiving, at the neural network, real or simulated multipath corrupted signal data, and training the neural network on the multipath corrupted signal data using a complex iterated least square thresholding algorithm (CILST) capable of processing both real and complex signals.

First claim

What is claimed is:

1 . A method for training a neural network to be configured to filter a multipath corrupted signal, said method comprising:
receiving, at the neural network, real or simulated multipath corrupted signal data; and training the neural network on the multipath corrupted signal data using a complex iterated least square thresholding algorithm (CILST) capable of processing both real and complex signals.

2 . A method in accordance with claim 1 , wherein receiving real or simulated multipath corrupted signal data comprises receiving real or simulated multipath corrupted signal data at a complex-valued neural network.

3 . A method in accordance with claim 1 , wherein receiving real or simulated multipath corrupted signal data comprises receiving real or simulated multipath corrupted signal data at a neural network that includes an input layer having a plurality of input nodes, a hidden layer having a plurality of hidden nodes, and an output layer having a plurality of output nodes.

4 . A method in accordance with claim 1 , wherein training the neural network using a CILST comprises:
for each of a plurality of hidden nodes of the neural network:
solving a complex linear least-square problem for the hidden node of the neural network to determine a linear fit to the multipath corrupted signal data;
computing an error from the linear fit at the hidden node; and
performing a new fit; and
fitting vector data for the plurality of hidden nodes by solving the complex linear least-square problem and computing a final real result.

5 . A method in accordance with claim 4 , wherein solving a complex linear least-square problem for the hidden node comprises solving the complex linear least-square problem for the hidden node in accordance with for t −1 (t −1 (ŷ))≈A 1 [{circumflex over (x)},1] for A 1 , where [•,1] denotes a matrix whose last column is all 1's, A is a complex matrix of weights from a complex input sample vector x, A j is a jth row of A, and t is an invertible threshold function that produces a positive real output, y, from a complex input.

6 . A method in accordance with claim 5 , wherein the invertible threshold function t is defined as

t


(
z
)

=



1

1
+



-
z



.

7 . A method in accordance with claim 5 , wherein computing an error comprises computing an error according to e 1 =t −1 (ŷ)−z 1 , where z 1 =min{|t(A 1 [{circumflex over (x)},1])|,1}.

8 . A method comprising:
receiving a multipath corrupted signal at a computing device; and processing the multipath corrupted signal using a neural network that is trained using a complex iterated least square thresholding algorithm (CILST) capable of processing both real and complex signals.

9 . A method in accordance with claim 8 , wherein processing the multipath corrupted signal comprises processing the multipath corrupted signal using a complex-valued neural network.

10 . A method in accordance with claim 8 , wherein processing the multipath corrupted signal comprises processing the multipath corrupted signal using a neural network that includes an input layer having a plurality of input nodes, a hidden layer having a plurality of hidden nodes, and an output layer having a plurality of output nodes.

11 . A method in accordance with claim 8 , wherein processing the multipath corrupted signal comprises processing the multipath corrupted signal using a neural network trained by:
for each of a plurality of hidden nodes of the neural network:
solving a complex linear least-square problem for the hidden node of the neural network to determine a linear fit to real or simulated multipath corrupted signal data;
computing an error from the linear fit at the hidden node; and
performing a new fit; and
fitting vector data for the plurality of hidden nodes by solving the complex linear least-square problem and computing a final real result.

12 . A method in accordance with claim 11 , wherein solving a complex linear least-square problem for the hidden node comprises solving the complex linear least-square problem for the hidden node in accordance with t −1 (t −1 (ŷ))≈A 1 [{circumflex over (x)},1] for A 1 , where [•,1] denotes a matrix whose last column is all 1's, A is a complex matrix of weights from a complex input sample vector x, A j is a jth row of A, and t is an invertible threshold function that produces a positive real output, y, from a complex input.

13 . A method in accordance with claim 12 , wherein the invertible threshold function t is defined as

t


(
z
)

=



1

1
+



-
z



.

14 . A method in accordance with claim 11 , wherein computing an error comprises computing an error according to e 1 =t −1 (ŷ)−z 1 , where z 1 =min{|t(A 1 [{circumflex over (x)},1])|,1}.

15 . A radar system configured to transmit and receive radar signals to determine a location of at least one object, said radar system comprising a computing device configured to:
receive a multipath corrupted radar signal; and process the multipath corrupted radar signal to generate a filtered radar signal, the multipath corrupted radar signal processed using a neural network that is trained using a complex iterated least square thresholding algorithm (CILST) capable of processing both real and complex signals.

16 . A radar system in accordance with claim 15 , wherein the neural network is a complex-valued neural network.

17 . A radar system in accordance with claim 15 , wherein the neural network includes an input layer having a plurality of input nodes, a hidden layer having a plurality of hidden nodes, and an output layer having a plurality of output nodes.

18 . A radar system in accordance with claim 15 , wherein the neural network is trained by:
for each of a plurality of hidden nodes of the neural network:
solving a complex linear least-square problem for the hidden node of the neural network to determine a linear fit to multipath corrupted signal data;
computing an error from the linear fit at the hidden node; and
performing a new fit; and
fitting vector data for the plurality of hidden nodes by solving the complex linear least-square problem and

Assignees

• Boeing Co

Inventors

• Ray Gary A

Classifications

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• H03H17/0248

  Filters characterised by a particular frequency response or filtering method · CPC title