Granted Patent B1

US 12,566,244 · App. 18/525,516 · Granted Mar 3, 2026

Inventors: Peter Petre (Oceanside, CA); Amir M. Rahimi (Santa Monica, CA)

Assignee: HRL LABORATORIES, LLC

G01S7/2886G01S13/89

Abstract

Claims (25)

1 . A system for real-time radar range-doppler map learning, the system comprising:

a physics enhanced dynamic reservoir core operable for receiving as an input signal a time delayed copy of a transmitted complex in-phase and quadrature (I/Q) radar waveform, and mapping the input signal onto a neural network to determine complex-valued output weights of neural states of the neural network;

an online learning layer that adapts the complex-valued output weights of the neural states to predict a most likely next value of the input signal, wherein the complex-valued output weights are continuously updated via a gradient descent online learning algorithm;

a neural combiner for combining a set of delayed neural state vectors with the complex-valued output weights of the online learning layer to compute a denoised and decluttered output signal, the output signal being separate in-phase and quadrature signals.

2 . The system as set forth in claim 1 , wherein the gradient descent online learning algorithm includes both a prediction error and L2 and L1 norm penalty terms.

3 . The system as set forth in claim 2 , wherein the L2 norm penalty term forces values of the complex-valued output weights to maintain within predetermined bounding limits.

4 . The system as set forth in claim 3 , wherein the L1 norm penalty term forces a learned weight matrix to a sparse solution that zeroes out noise generated false targets.

5 . A computer program product for real-time radar range-doppler map learning, the computer program product comprising:

a non-transitory computer-readable medium having executable instructions encoded thereon, such that upon execution of the instructions by one or more processors, the one or more processors perform operations of:

receiving as an input signal a time delayed copy of a transmitted complex in-phase and quadrature (I/Q) radar waveform;

mapping the input signal onto a neural network to determine complex-valued output weights of neural states of the neural network;

adapting the complex-valued output weights of the neural states to predict a most likely next value of the input signal, wherein the complex-valued output weights are continuously updated via a gradient descent online learning algorithm; and

combining a set of delayed neural state vectors with the complex-valued output weights of the online learning layer to compute a denoised and decluttered output signal, the output signal being separate in-phase and quadrature signals.

6 . The computer program product as set forth in claim 5 , wherein the gradient descent online learning algorithm includes both a prediction error and L2 and L1 norm penalty terms.

7 . The computer program product as set forth in claim 6 , wherein the L2 norm penalty term forces values of the complex-valued output weights to maintain within predetermined bounding limits.

8 . The computer program product as set forth in claim 6 , wherein the L1 norm penalty term forces a learned weight matrix to a sparse solution that zeroes out noise generated false targets.

9 . A computer implemented method for real-time radar range-doppler map learning, the method comprising an act of:

causing one or more processers to execute instructions encoded on a non-transitory computer-readable medium, such that upon execution, the one or more processors perform operations of:

receiving as an input signal a time delayed copy of a transmitted complex in-phase and quadrature (I/Q) radar waveform;

mapping the input signal onto a neural network to determine complex-valued output weights of neural states of the neural network;

adapting the complex-valued output weights of the neural states to predict a most likely next value of the input signal, wherein the complex-valued output weights are continuously updated via a gradient descent online learning algorithm; and

combining a set of delayed neural state vectors with the complex-valued output weights of the online learning layer to compute a denoised and decluttered output signal, the output signal being separate in-phase and quadrature signals.

10 . The method as set forth in claim 9 , wherein the gradient descent online learning algorithm includes both a prediction error and L2 and L1 norm penalty terms.

11 . The method as set forth in claim 10 , wherein the L2 norm penalty term forces values of the complex-valued output weights to maintain within predetermined bounding limits.

12 . The method as set forth in claim 10 , wherein the L1 norm penalty term forces a learned weight matrix to a sparse solution that zeroes out noise generated false targets.

Assignments (1)

Continuity (7)

References Cited (308)