System and method for linear prediction

• US 20020065664A1
• Filed: 10/10/2001
• Published: 05/30/2002
• Est. Priority Date: 10/13/2000
• Status: Active Grant

• Current Assignee: Leidos, Inc. (Leidos Holdings, Inc.)
• Original Assignee: Science Applications International Corporation
• Inventors: Goldstein, Jay Scott, Witzgall, Hanna Elizabeth
• Granted Patent: US 7,103,537 B2
• Time in Patent Office: Days
• US Class Current: 704/500
• CPC Class Codes: G06F 2218/22 Source localisation; Inverse modelling
• CPC Class Codes: G10L 19/04 using predictive techniques
• CPC Class Codes: G10L 25/12 the extracted parameters being prediction coefficients

Abstract

In a digital signal processor (DSP), input data is configured as a data matrix comprising data samples collected from an input signal. A weight vector is applied to the data matrix, where the weight vector comprises three parts including (a) a rank reduction transformation produced by decomposition of the data samples in a multistage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces. Each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage, to form a reduced rank data matrix. Part (b) of the weight vector comprises minimizing the mean squared error in the reduced rank data space. The output is a linear estimate of the input data. In one embodiment, further processing produces an estimate of the power spectral density of the input signal.

Claims

1. In a method for linear prediction of information determinable from at least one input signal comprising a plurality of data points, a method for training a linear prediction filter comprising:

providing at least one reference signal comprising reference data having known values;

collecting observed data corresponding to the at least one reference signal;

identifying a reduced order data space comprising a reduced order Krylov subspace between the observed data and the reference data; and

maximizing mutual data between the observed data and the reference data in the reduced order data space to define an autoregressive weight for use in the linear prediction filter.

2. The method of claim 1, wherein the step of identifying a reduced order data space comprises applying a multi-stage Wiener filter to the observed data.

3. The method of claim 1, wherein each of the at least one input signal and the at least one reference signal comprises a plurality of signals received at a multi-element detector array so that each element of the array provides a data point, and wherein the information comprises signal characteristics of the plurality of signals.

4. The method of claim 3, wherein the plurality of signals arrive at the array at a plurality of different angles and the information comprises angle of arrival.

5. The method of claim 3, wherein the reduced rank has a value that is much smaller than the number of elements in the array.

6. The method of claim 1, wherein the information comprises a frequency spectrum.

7. The method of claim 1, wherein the reduced rank is much smaller than the quantity of data points.

8. The method of claim 1, wherein the step of collecting observed data comprises collecting a plurality of snapshots of the at least one reference signal.

9. The method of claim 1, wherein the plurality of snapshots comprises time series data.

10. The method of claim 1, further comprising generating an error vector using the reduced rank data matrix.

11. The method of claim 10, wherein the error vector is generated by performing a Gram-Schmidt transformation on the reduced rank data matrix.

12. The method of claim 1, wherein the autoregressive weight is defined by minimizing the mean squared error.

13. The method of claim 12, wherein the information to be predicted comprises an estimated spectrum and further comprising using the autoregressive weight and the mean squared error to calculate the estimated spectrum.

15. The method of claim 14, wherein the step of processing the p observed data points to identify a reduced order data space comprises applying a multi-stage Wiener filter to the observed data.

16. The method of claim 14, wherein each of the at least one input signal and the at least one reference signal comprises a plurality of signals received at a multi-element detector array so that each element of the array provides a data point, and wherein the information comprises signal characteristics of the plurality of signals.

17. The method of claim 16, wherein the plurality of signals arrive at the array at a plurality of different angles and the information comprises angle of arrival.

18. The method of claim 16, wherein the reduced rank has a value that is much smaller than the number of elements in the array.

19. The method of claim 14, wherein the information comprises a frequency spectrum.

20. The method of claim 14, wherein the reduced rank is much smaller than the quantity of data points.

21. The method of claim 14 wherein collecting observed data comprises collecting a plurality of snapshots of the at least one reference signal.

22. The method of claim 14, wherein the plurality of snapshots comprises time series data.

24. The method of claim 23, wherein each of the at least one input signal and the at least one reference signal comprises a plurality of signals received at a multi-element detector array so that each element of the array provides a data point, and wherein the information comprises signal characteristics of the plurality of signals.

25. The method of claim 24, wherein the plurality of signals arrive at the array at a plurality of different angles and the information comprises angle of arrival.

26. The method of claim 24, wherein the reduced rank has a value that is much smaller than the number of elements in the array.

27. The method of claim 23, wherein the information comprises a frequency spectrum.

28. The method of claim 23 wherein the reduced rank is much smaller than the quantity of data points.

29. The method of claim 23, wherein the step of collecting observed data comprises collecting a plurality of snapshots of the at least one reference signal.

30. The method of claim 29, wherein the plurality of snapshots comprises time series data.

31. The method of claim 23, wherein the step of collecting received data comprises collecting a plurality of snapshots of the at least one received signal.

32. The method of claim 23, wherein the plurality of snapshots comprises time series data.

33. The method of claim 23, further comprising generating an error vector using the reduced rank data matrix.

34. The method of claim 33, where generating an error vector comprises performing a Gram-Schmidt transformation on the reduced rank data matrix.

35. The method of claim 23, wherein the information is an estimate of the spectrum of the at least one received signal.

37. The system of claim 36, further comprising a receiver for receiving the at least one signal, the receiver comprising a plurality of elements in an array, each element generating a data point of the plurality of data points for each of the plurality of data samples.

38. The system of claim 36, wherein the predicted signal comprises an estimated spectrum.

39. The system of claim 36, wherein the at least one signal comprises a plurality of signals impinging at a plurality of angles on a detector array comprising a plurality of elements, wherein the information comprises angle of arrival.

40. The system of claim 36 wherein the filter and the processor are incorporated in a digital signal processor.

14. A method for linear prediction of information determinable from at least one input signal comprising a plurality of data points, the method comprising:

providing at least one reference signal comprising a plurality of reference data points having known values;

training a filter by collecting p observed data points corresponding to p reference data points from the reference signal, wherein p reference data points comprise a portion of the plurality of reference data points, wherein training further comprises

processing the p observed data points through the filter to identify a reduced order data space comprising a reduced order Krylov subspace between the observed data points and the reference data points; and

defining a weight for minimizing the mean squared error between a predicted p+1 observed data point and a p+1 reference data point; and

applying the weight to filter the at least one input signal.

23. A method for linear prediction of information determinable from at least one signal received at a receiver adapted for receiving a set of data points, the at least one signal containing a plurality of data samples, the method comprising:

defining an observed data matrix comprising observed data samples collected from at least one reference signal received at the receiver, the at least one reference signal having a set of known data points;

defining a received data matrix comprising data samples collected from the at least one received signal;

applying a weight vector to the observed data matrix, the weight vector comprising;

(a) performing a rank reduction transformation produced by decomposition of the observed data matrix in a multi-stage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces, wherein a first stage comprises projecting the observed data matrix onto each of an initial first subspace comprising an initial normalized cross-correlation vector comprising a correlation vector between a known data point from the set of known data points and the observed data points in the receiver and an initial second subspace comprising the null space of the initial normalized cross-correlation vector, and each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage;

(b) minimizing the mean squared error in the reduced rank data matrix; and

applying the weight vector to the received data matrix.

36. A system for linear prediction of information determinable from at least one signal comprising a plurality of data points, the system comprising:

a linear prediction filter for processing a plurality of data samples collected from the at least one signal, the linear prediction filter comprising a multi-stage Wiener filter for projecting a full rank data matrix formed from the plurality of data samples and the plurality of data points into a subspace having a reduced rank to form a reduced rank data matrix and minimizing the mean squared prediction error in the reduced rank data space, and for applying a weight vector to the at least one signal to generate a predicted signal.

41. A method for spectrum estimation of at least one signal received at a detector comprising at least one detector element, the method comprising:

collecting a plurality of data snapshots from the at least one received signal;

defining a data matrix comprising the plurality of data snapshots;

applying a weight to the data matrix, the weight comprising;

(a) performing a rank reduction transformation produced by decomposition of the data matrix in a multi-stage Wiener filter having a plurality of stages, each stage comprising projection onto two subspaces, wherein a first stage comprises projecting the data matrix onto each of an initial first subspace comprising an initial normalized cross-correlation vector comprising a correlation vector between a known reference process and the remaining data points in the receiver and an initial second subspace comprising the null space of the initial normalized cross-correlation vector, and each subsequent stage comprises projecting data transformed by the preceding second subspace onto each of a first subspace comprising a normalized cross-correlation vector at the previous stage and a second subspace comprising the null space of the normalized cross-correlation vector of the current stage;

(b) minimizing the mean squared error in the reduced rank data matrix; and

using the weight and the mean squared error to define an autoregressive power spectrum.

42. The method of claim 41, wherein the at least one received signal comprises a plurality of signals received by a detector array having a plurality of detector elements wherein one detector element receives one data point for each of the plurality of data samples.

43. The method of claim 42, wherein the reduced rank has a value that is much smaller than the number of elements in the array.

44. The method of claim 41, wherein the plurality of signals arrive at the array at a plurality of different angles and the information comprises angle of arrival.

45. The method of claim 41, wherein the step of minimizing mean squared error comprises generating an error vector by performing a Gram-Schmidt transformation on the reduced rank data matrix.

46. The method of claim 41, wherein the reduced rank is much smaller than the quantity of data points.

47. The method of claim 41, wherein the step of collecting data samples comprises collecting a plurality of snapshots of the at least one signal.

48. The method of claim 41, further comprising generating an error vector using the reduced rank data matrix.

49. The method of claim 48, where generating an error vector comprises performing a Gram-Schmidt transformation on the reduced rank data matrix.

50. A method for spectrum estimation in at least one input signal received at a detector comprising at least one detector element, the method comprising:

collecting a plurality of snapshots of the at least one input signal, each snapshot comprising a plurality of observed data points;

defining a observed data matrix comprising the plurality of snapshots;

identifying a reduced order data space comprising a reduced order Krylov subspace between the observed data points and reference data from a known reference signal;

maximizing mutual data points between the observed data points and the reference data in the reduced order data space to define an autoregressive weight; and

using the autoregressive weight to calculate an estimated spectrum.

51. The method of claim 50, wherein the autoregressive weight is defined by minimizing the mean squared error.

52. The method of claim 50, wherein the at least input signals comprises a plurality of signals arriving at the array at a plurality of different angles and the estimated spectrum includes an angle of arrival for each of the plurality of signals.

53. A system for spectrum estimation in at least one input signal comprising a plurality of data points received at a detector comprising at least one detector element, the system comprising:

a filter for processing a plurality of data samples collected from the at least one signal, the filter comprising a multi-stage Wiener filter for projecting a full rank data matrix formed from the plurality of data samples and the plurality of data points into a subspace having a reduced rank to form a reduced rank data matrix and determining a weight for minimizing the mean squared error in the reduced rank data space; and

a processor for calculating an estimated spectrum using the weight and the mean squared error.

54. The system of claim 53, further comprising a receiver for receiving the at least one signal, the receiver comprising a plurality of elements in an array, each element generating a data point of the plurality of data points for each of the plurality of data samples.

55. The system of claim 53 wherein the filter and the processor are incorporated in a digital signal processor.