REVEAL Receiver Exploiting Variability in Estimated Acoustic Levels Project Review 16 Sept 2008 Presented to Program Officers: Drs. John Tague and Keith Davidson Undersea Signal Processing Team, Office of Naval Research Penn State Team: Faculty: R. Lee Culver 1 and Nirmal K. Bose 2 Students: Colin W. Jemmott 3, Jeremy Joseph 3, Brett Bissinger 2, and Alex Sell 3 1 Applied Research Lab, 2 Department of Electrical Engineering, and 3 Penn State University, State College, PA Contact info: 814 865-3383 rlc5@psu.edu 16 September 2008 REVEAL Overview and Progress 1
REVEAL Long Range Goals Develop a signal processing structure that exploits environmental knowledge by incorporating signal and noise predictions. Use this SP structure to develop improved detectors and classifiers which remain robust to variable and random signal and noise. No specific system application, but focus on passive sonar and frequency 1 khz. Train the future generation of ocean acousticians and signal processors. 16 September 2008 REVEAL Overview and Progress 2
REVEAL Project focus Since FY05, the project goal has been to work at the interface between OA and SP in order to apply and transition OA products to SP algorithms. Ocean acoustics REVEAL focus Sonar signal processing transition 16 September 2008 REVEAL Overview and Progress 3
REVEAL approach Ocean acoustic models and knowledge Compute signal and noise parameter statistics (also called prior statistics or training data) Passive beamformed sonar data M-ary detector or classifier decision Estimated Ocean Detector (composite Likelihood Ratio) Kullback-Leibler divergence (et. al.) Bayesian (histogram) filter 16 September 2008 REVEAL Overview and Progress 4
Classification using discriminant functions Signal parameters Discriminant functions 16 September 2008 REVEAL Overview and Progress 5
Typical problem The signal is affected by propagation through the ocean, and we have knowledge and models for the oceanic properties and processes that affect acoustic propagation. Our approach is to use Monte Carlo simulation to obtain many realizations of the signal from statistically-valid realizations of the environment in order to classify the signal source. p 1 (s(θ)) pdf of signal from near-surface source (H 1 ) Source near the surface Receive array p 2 (s(θ)) pdf of signal from near-bottom source (H 2 ) Source near the bottom 16 September 2008 REVEAL Overview and Progress 6
Composite LR Consider observation r = s( θ ) + n. A composite Likelihood Ratio (LR) incorporates statistical knowledge of random parameter θ : ( r) ( H1) ( ) p r Λ = = [ H, ] ( H ) pr θ pθ dθ 1 1 H 2 H 2, H 2 [ ] ( ) p r pr θ pθ dθ Since the noise is additive, the likelihood function is the pdf of the noise: ( ) [ θ] = ( θ) θ = ( θ) pr H, pr s H, p r s H If the noise is Gaussian, the likelihood function is then : 1 1 n 1 1 pn ( r s( θ) H, θ) = exp r s ( θ ) 1 1 2 2 2σ 2σ 1 1 16 September 2008 REVEAL Overview and Progress 7 2
Composite LR (cont) and the Likelihood Ratio (LR) is : 2 ( r s1 ) exp 2 p( θ H1 ) dθ 2σ 1 Λ ( r) = 2 ( r s2 ) exp p 2 ( θ H2 ) dθ 2σ 2 The Estimator-Correlator (EC) provides an expression for the LR in the more general case where the noise pdf belongs to the exponential class. Jeff Ballard formulated the EC for Gaussian signals in FY07 and sinusoids in FY08. 1. Schwartz, S. C., The Estimator-Correlator for Discrete-Time Problems, IEEE Trans. on Inf. Theory, Vol. 23, No. 1, Jan 1977, pp. 93-100. 2. Ballard and Culver, The Estimated Signal Parameter Detector: Incorporating signal parameter statistics in the signal processor, submitted to JOE (2008). 16 September 2008 REVEAL Overview and Progress 8
Signal Parameter pdf ( 1 ( θ) ) p pθ H 1 Signal Parameter pdf Estimated Ocean Detector (EOD) p Conditional Moment Function Received Signal, r ( r θ ) 1( ) h r h 2( r ) G G B 1 2 1 B 2 Noise pdf ( r ) ( r ) ( r ) ( r ) ln c 1 c 2 - H Σ ( ) > 1 r - Noise only data < ln η H 2 H 1 H 2 ( p ( θ )) p θ H 2 16 2September 2008 REVEAL Overview and Progress 9
Neglecting noise When the noise is neglibly small, the likelihood function becomes ( r) ( H, θ) = ( θ) H, θ δ ( θ) p r 1 p r s 1 r s and the Likelihood Ratio is then p r s( θ) H 1, θ p s( θ) H 1 dθ p s( θ ) H 1 p r s( θ) H 2, θ p s( θ) H 2 dθ p s ( θ ) H2 Λ = We have made this assumption in applying the composite LR to the 1996 Strait of Gibraltar and Swellex-96 data, respectively. 1. Culver, R. L. and H. J. Camin, Dependence of probabilistic acoustic signal models on statistical ocean environmental models, submitted to JASA (2008). 2. Jemmott, C.W., R. L. Culver, and N. K. Bose, Passive sonar depth classification using model based amplitude statistics, (in preparation). 16 September 2008 REVEAL Overview and Progress 10
The Bayes Filter The Bayes filter is an alternative to the LR in which we use Bayes rule ( H, θ) ( H θ) = (,H θ) = ( H, θ) ( θ) p r p p r p r p r i i i i to convert the likelihood function to the posterior pdf ( θ,h i) p( θ Hi) p( r θ ) p r = ( r θ ) H,. We select the hypothesis with the highest posterior probability. The histogram filter is the discrete implementation of the Bayesian filter. Colin will compare a recursive histogram filter to the LR receiver. p i 16 September 2008 REVEAL Overview and Progress 11
Distance measures The Kullback-Leibler divergence (among others) provides a measure of the distance between two multidimensional surfaces, e.g. pdfs. Using a distance measure to classify signals: Predict signal parameter pdfs for difference classes Estimate signal parameters from observations; compute signal parameter pdfs from observations Pick the class whose pdf is closest to the observed signal parameter pdf Brett will present his work on this approach 16 September 2008 REVEAL Overview and Progress 12
Noise whitening The EOD requires that the noise pdf belong to the exponential class. Not necessarily Gaussian. How to whiten or decorrelate? So-called higher order whitening has been investigated in the image processing literature. Whitening is closely related to distance measures and to compressive sampling. Dr. Bose will present his work. 1. J. Gluckman, Higher order whitening of natural images, Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2005. 2. N.K. Bose, U. Srinivas and R.L. Culver (2008). Wavelength diversity based infrared super-resolution and condition-based maintenance, INSIGHT, British Institute for Non-Destructive Testing, Vol. 50, No. 8. 16 September 2008 REVEAL Overview and Progress 13
Predicting signal parameter pdfs Rough surface PE acoustic propagation model obtained from Rosenberg (APL/JHU); based upon Range-dependent Acoustic Model (RAM), but adds capability for acoustic propagation with time- and spatially-varying rough surface. We want to determine if this simulation can predict surface interaction effects on signal frequency and amplitude pdfs. Jeremy Joseph will present his work. 16 September 2008 REVEAL Overview and Progress 14
REVEAL FY08 Talks Acoust. Soc. Am. Nov 07 (New Orleans) Culver: Likelihood func. & signal param. pdfs for sonar signal processing Joseph: Effect of rough surface on received signals Apr 08 PSU College of Engr. Research Symposium (CERS) Jemmott: Passive sonar source classification based on received signal amplitude variation statistics Acoust. Soc. Am. June - July 08 (Paris) Bissinger: Application of statistical methods in uw signal classification Joseph: Effects of volume and boundary variability on the statistics of received signal frequency Culver: Detection and classification using the Estimated Ocean Detector Distributed Detection and Est. Workshop (VA Tech, July 08) Bissinger: Statistical Distance Based Signal Classification Culver: Detection and classification using the Estimated Ocean Detector Jemmott: Passive Sonar Model-Based Source Location Classification Joseph: Effect of rough surface on received signals 16 September 2008 REVEAL Overview and Progress 15
REVEAL FY08 Papers N.K. Bose, U. Srinivas and R.L. Culver (2008). Wavelength diversity based infrared super-resolution and condition-based maintenance, INSIGHT, British Institute for Non-Destructive Testing, Vol. 50, No. 8. J.A. Ballard and R. L. Culver (2008). The Estimated Signal Parameter Detector: Incorporating signal parameter statistics into the signal processor submitted to IEEE J. Oceanographic Engr. Dec 07 Comments received June 08 Manuscript revised and resubmitted Sept 08 R.L. Culver and H.J. Camin (Dependence of probabilistic acoustic signal models on statistical ocean environmental models submitted to J. Acoust. Soc. Am. Nov 07 Comments received May 08 Manuscript under revision; will re-submit by 30 Sep 08 16 September 2008 REVEAL Overview and Progress 16
Planned FY09 Talks and Papers Present at Asilomar (Oct 08), CISS (Mar 09), and UASP (Oct 09) signal processing meetings. Present as ASA (Nov 09) and (May 09) Publish papers on: depth classification using the Swellex-96 data higher order whitening application of distance measure to uw acoustics predicting uw acoustic signal parameter statistics 16 September 2008 REVEAL Overview and Progress 17
REVEAL Project FY09 Plans Apply classifiers to South Florida range data Bottom-mounted line arrays Surface ships and a towed, submerged source Move from high SNR to moderate SNR cases Incorporate noise whitening for EOD Robustness of distance measures and histogram filter Address correlation in extracted parameter values Incorporate the rough surface RAM simulation into the signal processing architecture 16 September 2008 REVEAL Overview and Progress 18
REVEAL Project Ocean environment ocean models In-situ measurements environ. realizations acoustic propagation model Monte Carlo simulation signal realizations sonar data noise pdf Classifier Detection/classification decision Signal parameter pdf s (MaxEnt method) 16 September 2008 REVEAL Overview and Progress 19
REVEAL Project Ocean environment ocean models In-situ measurements environ. realizations acoustic propagation model Monte Carlo simulation signal realizations sonar data noise pdf Colin and Brett Classifier Detection/classification decision Signal parameter pdf s (MaxEnt method) 16 September 2008 REVEAL Overview and Progress 20
REVEAL Project Ocean environment ocean models In-situ measurements environ. realizations acoustic propagation model Monte Carlo simulation signal realizations sonar data noise pdf Dr. Bose Classifier Detection/classification decision Signal parameter pdf s (MaxEnt method) 16 September 2008 REVEAL Overview and Progress 21
REVEAL Project Ocean environment ocean models In-situ measurements environ. realizations Jeremy acoustic propagation model Monte Carlo simulation signal realizations sonar data noise pdf Classifier Detection/classification decision Signal parameter pdf s (MaxEnt method) 16 September 2008 REVEAL Overview and Progress 22