Method for detecting weak signals in a non-gaussian and non-stationary background

Abstract

A method is described for detecting weak signals in a non-Gaussian and non-stationary background using a hidden Markov parameter estimator. The method comprises the steps of: a) partitioning input data into sets of range bins, where the input data has a noise component; b) estimating hidden Markov model parameters from the input data; c) determining the average intensity I.sub.i of the range bins, where i represents an index for the ranges bins, and 1.ltoreq.i .ltoreq.R, and R represents the total number of range bins; d) estimating the average noise intensity N.sub.i of each range bin; e) determining the residual intensity estimates .rho..sub.i of each range bin as the maximum of 0 and (I.sub.i -N.sub.i) for all values of i; f) transforming the hidden Markov model parameters into estimates of hidden Markov model parameters of the noise component of the input data; g) generating modified locally optimal detection statistics representing the likelihood that the input data for the range bins contains a signal of interest using the estimates of the hidden Markov model parameters of the noise component of the input data, the residual intensity estimates, and the input data; h) transforming the modified locally optimal detection statistic into normalized values of the detection statistics applied to each range bin; and i) generating a detection output signal if any of the normalized values of the detection statistics exceeds a threshold.

Description

BACKGROUND OF THE INVENTION

The present invention relates to the field of signal processing, and more particularly, to a method for detecting weak signals in a non-stationary, non-Gaussian background using a hidden Markov parameter estimator.

Radar systems are often used to detect signals from non-stationary and non-Gaussian environments. One type of system for detecting signals from such environments is described with reference to U.S. Pat. No. 5,694,342, entitled A Method for Detecting Signals In Non-Gaussian Background Clutter. In the system described in the '342 patent, input data represents a collection of successive intensities (norm squared) of baseband demodulated range-walk-corrected radar returns from a set of range bins organized into a range-by-pulse matrix X=(x.sub.ij), where R is the number of range bins, P is the number of pulses, and 1.ltoreq.i.ltoreq.R, 1.ltoreq.j.ltoreq.P. The data are filtered to partition range bins having exponentially distributed data from those that have non-exponentially distributed data. The intensities of the exponentially distributed data are estimated. Exponential mixture distributions are fit to each range bin of the non-exponential data. Then, noise parameters are selected for each range bin. The residual intensity of the data in each range bin is estimated. A detection statistic M.sub.i and the standard deviation N.sub.i are determined for each range bin. A normalized detection statistic S.sub.i is defined by S.sub.i =M.sub.i /N.sub.i. The maximum value, S.sub.max, and the mean, S.sub.mean, and standard deviation, S.sub.std, of all S.sub.i excluding S.sub.max are determined. A threshold .tau..sub..alpha. corresponding to a false alarm probability .alpha. is determined. An output signal is generated for range bin i if (S.sub.i -S.sub.max)/S.sub.std .gtoreq..tau..sub..alpha..

The system described in the '342 patent generally requires 50 or more data samples, and more preferably, 100+ data samples, to perform reasonable estimates of the model parameters. It would be desirable to develop a system that could detect weak signals in a non-Gaussian, non-stationary background which required fewer data samples.

SUMMARY OF THE INVENTION

A method is described for detecting weak signals in a non-Gaussian and non-stationary background using a hidden Markov parameter estimator. The method comprises the steps of: a) partitioning input data into sets of range bins, where the input data has a noise component; b) estimating hidden Markov model parameters from the input data; c) determining the average intensity I.sub.i of the range bins, where i represents an index for the ranges bins, and 1.ltoreq.i.ltoreq.R, and R represents the total number of range bins; d) estimating the average noise intensity N.sub.i of each range bin; e) determining the residual intensity estimates .rho..sub.i of each range bin as the maximum of 0 and (I.sub.i -N.sub.i) for all values of i; f) transforming the hidden Markov model parameters into estimates of hidden Markov model parameters of the noise component of the input data; g) generating modified locally optimal detection statistics representing the likelihood that the input data for the range bins contains a signal of interest using the estimates of the hidden Markov model parameters of the noise component of the input data, the residual intensity estimates, and the input data; h) transforming the modified locally optimal detection statistic into normalized values of the detection statistics applied to each range bin; and i) generating a detection output signal if any of the normalized values of the detection statistics exceeds a threshold.

The invention preferably may be implemented as a set of computer readable program instructions which may be encoded onto a program storage device such as magnetic tape, a floppy disk, magnetoptical disk, compact disk, computer memory, or the like. The storage device embodies the program of instructions for implementing the function of detecting weak signals in a non-Gaussian and non-stationary backgrounds.

The hidden Markov parameter estimator offers improved performance and simpler implementations compared to the prior art exponential mixture detector. The hidden Markov parameter estimator generalizes the exponential mixture model and may more accurately represent backgrounds such as radar sea clutter. The correlation structure of the hidden Markov parameter estimator provides a simpler way to treat the non-stationary components of the data as time series having a variety of probability distributions. The invention estimates the Markov model parameters on blocks of range bins which leads to more accurate estimates of the noise parameters for the input data.

These and other advantages of the invention will become more apparent upon review of the accompanying drawings and specification, including the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of a system for detecting signals in non-Gaussian and non-stationary background which embodies various features of the present invention.

FIG. 2 is a graph which shows range data, X.sub.ij, associated with a particular pulse having an index j, and a particular distance associated with an index value, i, where i and j are positive integers.

DESCRIPTION OF THE PREFERRED EMBODIMENT

The following description is of the best mode presently contemplated for carrying out the invention. This description is not to be taken in a limiting sense, but is made merely for the purpose of describing the general principles of the invention. The scope of the invention should be determined with reference to the claims.

Referring to FIG. 1, the present invention employs a hidden Markov parameter estimator for providing a method 10 for detecting weak signals in a non-Gaussian and non-stationary background such as radar sea clutter, acoustic data containing moving interferers, and seismic data. The invention preferably may be implemented as computer readable program instructions which may be encoded onto a program storage device such as magnetic tape, a floppy disk, magnetooptical disk, compact disk, computer memory, or the like. The storage device embodies the program of computer executable instructions for implementing the functions of the detector.

Method 10 comprises a hidden Markov model parameter estimator 14, parameter smoother 16, residual intensity estimator 18, detection statistic calculator 20, adaptive normalizer 24, and a threshold comparator 28. ATTACHMENT 1 provides, by way of example, a listing of a computer program for implementing the invention, and more particularly, each of the elements depicted in FIG. 1. A cross-reference between each of the functional blocks in FIG. 1 and specific program listings is provided in TABLE 1, below. The computer program represented in ATTACHMENT 1 was written in MATLAB.RTM.. However, it is to be understood that software for implementing the invention may be written in any suitable computer language.

TABLE 1______________________________________CROSS REFERENCE BETWEEN FUNCTIONAL BLOCKS OF FIG. 1AND PROGRAM LISTINGS IN ATTACHMENT 1Function Depicted In FIG. 1 Lines In Attachment 1______________________________________HMM Parameter Estimator 14 211-306Parameter Smoother 16 128-167Residual Intensity Estimator 18 489-492 and 536-600Detection Statistic Calculator 20 603-638Adaptive Normalizer 24 172-185Threshold Comparator 28 186, 187______________________________________

Input data 12 may represent a collection of successive intensities (norm squared) of baseband demodulated optical, radar, or sonar signal returns comprising a set of range bins organized into a range-by-pulse matrix X=(x.sub.ij), where R is the number of range bins, P is the number of pulses and 1.ltoreq.i.ltoreq.R, 1.ltoreq.j.ltoreq.P. Input data may represent, by way of example, radar, sonar, or optical return signals. The return signals may also include electromagnetic radiation other than radar or optical signal returns. For example, if input data 12 represents radar range bin data consisting of 512 range bins and 32 pulses, input data 12 may be structured in a 512.times.32 matrix. Input data 20 are input to Hidden Markov Model (HMM) parameter estimator 14, detection statistic calculator 20, and residual intensity estimator 18. HMM parameter estimator 14 partitions the range bins into sub-blocks representing some number of range bins and generates HMM parameters 15 for each sub-block of range bins. Parameter smoother 16 is an averaging filter that transforms HMM parameters 15 into an output signal 17 representing estimates of the HMM parameters of the noise component of input data 20.

Residual intensity estimator 18 transforms input signal 12 into output signal 19 which represents estimates of the noise power for each range bin of input data 12. In response to receiving input signals 12, 17, and 19, detection statistic calculator generates a modified locally optimal detection statistic 22 representing the likelihood that the input data for a given range bin contains a signal of interest. Adaptive normalizer 24 transforms signal 22 into an output signal 26 representing normalized values of the detection statistic applied to each range bin. Adaptive normalizer 24 transforms the detection statistic values represented in output signal 22 for each range bin to units of standard deviation from the mean value of the detection statistic for that block of range bins. The range bins selected to compute the normalized output values in signal 26 are range bins in a sub-block except the range bin in question and a specified number of guard bands on either side of the range bin in question. Threshold comparator 28 compares the value of the normalized detection statistic from signal 26 with a predetermined threshold, and generates a detection output signal 30 for the range bin if the normalized value of signal 26 exceeds the threshold.

HMM parameter estimator 14 partitions input data 12 into blocks of range bins. Range bins are described with reference to FIG. 2. FIG. 2 is a graph which shows that various values of range data, X.sub.ij, are associated with a particular pulse having an index j, indicated on the horizontal axis, and a particular distance associated with an index value, i, shown on the vertical axis, where i and j are positive integers, P represents the total number of pulses, and Z represents the furthest distance for which range data is stored. In the case in which pulse number 1 has an index j where j=1, range data corresponding to signal returns are represented by range data X.sub.(i,j), where i=1 to Z. Range data, X.sub.(6,1) may represent a signal return from a distance associated with an index value equal to 1, where j=1. By way of example, pulse number 1 may be an acoustic, radar, or optical pulse. Range bins are defined as the set of range data X.sub.(c.sub.,j), where C is a positive integer constant. Contiguous range bins are defined as the set of range data X.sub.(a.sub.,j) and the set of range data X.sub.(.alpha..+-.1,j) where .alpha. is an integer constant, and 2.ltoreq..alpha..ltoreq.(z-1). Hidden Markov model parameters are estimated for each block of range data X.sub.(i,j), where 1.ltoreq.j.ltoreq.P, by using the expectation and maximization algorithm (EM). The number of states of the Markov model is denoted by m. For m=1, the data are modeled as having an exponential distribution, and the expected value of the data is estimated. A goodness of fit test, such as chi-squared, is used to compare the empirical distribution of the data and the model distribution. If the test indicates that the data do not have an exponential distribution, then hidden Markov model parameters are estimated for m=2. If the goodness of fit test fails for m=2, then the parameters are estimated for m=3, and so on. For m.gtoreq.2 the HMM is defined by the state transition parameters, the initial distribution of the states, and the distribution of the observations for given states. For 1.ltoreq.i, j.ltoreq.m, the probability of transiting from state i to state j is denoted by .alpha..sub.ij. The data generated by state j are assumed to have an exponential distribution with parameter .nu..sub.j, i.e.

p(x.vertline.j)=exp(-x/(2.nu..sub.j)). The initial probability of observing a datum from state j is .pi..sub.j.

The estimation of the values of the parameters {.alpha..sub.ij,.nu..sub.j,.pi..sub.j }.sub.1.ltoreq.i,j.ltoreq.m proceeds by 1) using the EM algorithm to estimate initial values of .nu..sub.j and .pi..sub.j for the stationary distribution of the HMM, which is an exponential distribution, selecting initial values for the .alpha..sub.ij, and applying the EM algorithm to obtain maximum likelihood estimates of the HMM parameters, or 2) by using values of the HMM parameters obtained from spatially-temporally continuous range as initial values of the EM algorithm and applying the EM algorithm.

The HMM parameters 15 are estimated from input data 12 that may contain signal as well clutter and noise, however detection statistic calculator 20 performs best using HMM parameters of the noise component of input data 12. Parameter smoother 16 reduces the effect of the possible presence of signal components on the HMM parameter estimates 15. HMM parameter estimates 15 may be represented by {.nu..sub.j,.alpha..sub.ij,.pi..sub.j }, for 1.ltoreq.ij.ltoreq.m, where m is the number of states, .nu..sub.j is the expected intensity of state j, .nu..sub.i <.nu..sub.j if i<j, .alpha..sub.ij is the probability of transitioning from state i to state j, and .pi..sub.j is the initial probability of state j. Parameter smoother 16 first removes the influence of a possible signal component from .nu..sub.1 by replacing the original estimate of .nu..sub.1 obtained for each block of range bins, with an order statistic applied to the set of original estimates of .nu..sub.1. If for each block of range bins the replacement value of .nu..sub.1 is below the original value, and m=1, then a second state is added by setting .nu..sub.2 equal to the original value of .nu..sub.1 and defining state transition probabilities .alpha..sub.ij for 1.ltoreq.i,j.ltoreq.2 and initial probabilities .pi..sub.j for 1.ltoreq.j.ltoreq.2. Next, parameter smoother 16 imposes a lower limit on the ratio .nu..sub.2 /.nu..sub.1. If .nu..sub.2 /.nu..sub.1 is below this limit, then .nu..sub.2 is replaced with a values such that the lower bound of the ratio .nu..sub.2 /.nu..sub.1 is achieved. By way of example, the lower limit may be selected whereby the lower limit of .nu..sub.2 /.nu..sub.i is equal to 1.5, a value developed empirically. However, it is to be understood that other values for the lower limit of .nu..sub.2 /.nu..sub.i may also be used to suit the requirements of a particular application. The third major step performed by parameter smoother 16 is to place an upper bound on .alpha..sub.11 by replacing any values of .alpha..sub.11 that exceed the value of the upper bound with the value of the upper bound and adjust other transition probabilities to maintain the relationship: ##EQU1##

Residual intensity estimator 18: 1) determines the average intensity of each range bin; 2) estimates the noise power for each range bin; and 3) substracts the noise power estimate from the value associated with the average intensity of each range bin if the noise power estimate is below the average intensity of the range bin. This difference is provided as output signal 19 and is denoted .rho..sub.i. The noise power estimates may be obtained by using order statistics to obtain initial estimates and a linear prediction filter to estimate local maxima of the noise power intensity that are not well represented by the order statistics. If two estimates of the noise power are determined for a range bin, then residual intensity estimator 18 selects the larger value representing the noise power. b.sub.i (o.sub.j)

Detection statistic calculator 20 calculates a detection statistic for each range bin. The detection statistic is; defined in terms of input data 12, HMM parameters 15, and the scaled forward and backward coefficients, .alpha..sub.j (i) and .beta..sub.j (i), respectively of HMM parameters 15, as defined in equations 2 through 14, modified to include the residual intensity determined by adding .rho..sub.i and .nu..sub.i in the definition of b.sub.i (o.sub.j) defined further herein. The output values 22 of detection statistic calculator 20 are: ##EQU2## The scaled forward and backward coefficients of HMM parameter estimates 15, .alpha..sub.j (i) and .beta..sub.j (i), respectively are defined as follows. Assume that HMM parameters {.alpha..sub.ij,.nu..sub.j,.pi..sub.j } have been obtained for observations O=(o.sub.1, . . . ,o.sub.n), and that b.sub.i (o.sub.j) is the probability of o.sub.j given state i, i.e: ##EQU3## Then the scaled forward coefficients are defined recursively by ##EQU4## The scaled backwards coefficients are also defined recursively by: ##EQU5##

Adaptive normalizer 24 transforms signal 22 into an output signal 26 representing normalized values of the detection statistic applied to each range bin. For each range bin, the mean S.sub.mean and standard deviation S.sub.std of the output values 22 provided by detection statistic calculator 20 from the bins of the block excluding the given bin and guard bins are calculated. Output signal 22 of detection statistic calculator 20 on the test bin is normalized as T=(LO(O)-S.sub.mean)/S.sub.std, where T represents the output value 26 from adaptive normalizer 24.

Threshold comparator 28 compares the normalized output values 26 from adaptive normalizer 24 to a threshold .tau.. Threshold comparator 28 generates an output signal 30 is generated if T>.tau..

Obviously, many modifications and variations of the present invention are possible in light of the above teachings. It is therefore to be understood that within thescope of the appended claims, the invention may be practiced otherwise than as specifically described. ##SPC1##

Claims

1. A set of computer readable program instructions for implementing a method for detecting weak signals in a non-Gaussian and non-stationary background, comprising the steps of:

a) partitioning input data into sets of range bins, where said input data has a noise component;

b) estimating hidden Markov model parameters from said input data;

c) determining an average intensity I.sub.i of said range bins, where i represents an index for said ranges bins, 1.ltoreq.i.ltoreq.R0, and R represents the total number of range bins;

d) estimating an average noise intensity N.sub.i of each said range bin;

e) determining a residual intensity estimates .rho..sub.i of each said range bin as the maximum of 0 and (I.sub.i -N.sub.i) for all values of i;

f) transforming said hidden Markov model parameters into estimates of hidden Markov model parameters of said noise component of said input data;

g) generating modified locally optimal detection statistics representing the likelihood that said input data for said range bins contains a signal of interest using said estimates of said hidden Markov model parameters of said noise component of said input data, said residual intensity estimates, and said input data;

h) transforming said modified locally optimal detection statistic into normalized values of said detection statistics applied to each said range bin; and

i) generating a detection output signal if any of said normalized values of said detection statistics exceeds a threshold.

2. The method of claim 1 wherein step (a) further includes partitioning said input data into sets of contiguous range bins.

3. The method of claim 1 wherein said input data respresents signal returns.

4. The method of claim 3 wherein said signal returns include clutter.

5. The method of claim 3 wherein said signal returns are selected from the group consisting of radar signal returns, acoustic energy signal returns, and optical energy signal returns.

6. A program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine to perform method steps, the method steps comprising:

a) partitioning input data into sets of range bins, where said input data has a noise component;

b) estimating hidden Markov model parameters from said input data;

c) determining an average intensity I.sub.i of said range bins, where i represents an index for said ranges bins, and 1.ltoreq.i.ltoreq.R, and R represents the total number of range bins;

d) estimating an average noise intensity N.sub.i of each said range bin;

e) determining a residual intensity estimates .rho..sub.i of each said range bin as the maximum of 0 and (I.sub.i -N.sub.i) for all values of i;

f) transforming said hidden Markov model parameters into estimates of hidden Markov model parameters of said noise component of said input data;

g) generating modified locally optimal detection statistics representing the likelihood that said input data for said range bins contain a signal of interest using said estimates of said hidden Markov model parameters of said noise component of said input data, said residual intensity estimates, and said input data;

h) transforming said modified locally optimal detection statistic into normalized values of said detection statistics applied to each said range bin; and

i) generating a detection output signal if any of said normalized values of said detection statistics exceeds a threshold.

7. The method of claim 6 wherein step (a) further includes partitioning said input data into sets of contiguous range bins.

8. The method of claim 6 wherein said input data respresents signal returns.

9. The method of claim 8 wherein said signal returns include clutter.