You are on page 1of 195
PASSIVE LOCALIZATION OF AN UNDERWATER ACOUSTIC SOURCE USING DIRECTIONAL SENSORS A Thesis Submitted to the Faculty of Graduate Studies and Research in Partial Fulfilment of the Requirements for the Degree of Doctor of Philosophy in the Department of Electrical Engineering University of Saskatchewan by Yun Xiang Yuan Saskatoon, Saskatchewan October 1994 ©Copyright Y.X. Yuan, 1994, All rights reserved. National Library of Canada Acquisitions and Bibliographic Services 305 Welington Strect Qeawa ON KIA ONE Canada Bibliotheque nationale du Canada Acquisitions et ‘services bibliographiques 395, re Welington ‘Onawa ON KIA ONE Canada The author has granted a non- L’auteur a accordé une licence non exclusive licence allowing the exclusive permettant ala National Library of Canada to Bibliothéque nationale du Canada de reproduce, loan, distribute or sell _reproduire, préter, distribuer ou copies of this thesis in microform, vendre des copies de cette these sous paper or electronic formats la forme de microfiche/film, de reproduction sur papier ou sur format électronique. The author retains ownership of the _L’ auteur conserve la propriété du copyright in this thesis. Neither the droit d’auteur qui protege cette thése thesis nor substantial extracts from it Ni la thése ni des extraits substantiels may be printed or otherwise de celle-ci ne doivent étre imprimés reproduced without the author’s ou autrement reproduits sans son permission. autorisation. Canad 0-612-23926-8 UNIVERSITY OF SASKATCHEWAN College of Graduate Studies and Research SUMMARY OF DISSERTATION Submitted in partial fulfilment of the requirements for the DEGREE OF DOCTOR OF PHILOSOPHY by Yun Xiang Yuan Department of Electrical Engineering University of Saskatchewan Fall, 1994 Examining Committee: Dr. S. Sokhansanj Dr. R. J. Bolton Dr. J. E. Salt Dr. G. C. Carter Dr. L. G. Watson Dr. B. L. F. Daku Dr. R. E. Gander Dean/Associate Deaa/Dean’s Designate, Chair, College of Graduate Studies and Research Graduate Chair, Department of Electrical Engineering Co-supervisor, Department of Electrical Engineering Co-supervisor, NUWC, New London, CT, USA Department of Mechanical Engineering Department of Electrical Engineering Department of Electrical Engineering External Examiner: Dr. R. L. Kirlin Department of Electrical and Computer Engineering University of Victoria Victoria, BC VOW 2¥2 Canada Passive Localization of an Underwater Acoustic Source using Directional Sensors ‘The accuracy of passive localization of a single stationary source is studied in an un- derwater acoustic environment. The bearing, range and depth of the source is estimated using a two-receiver vertical array in a direct and surface-reflected path environment. Each of the two receivers consists of a cluster of directional sensors uniformly located on a small circle. The gain profile of the directional sensors is such that it depends only on the source bearing. This novel receiver is proposed in this thesis to facilitate bearing estimation The source bearing is estimated from the average received signal power measurements using one of the two clusters. The feasibility of energy-based bearing estimators is investi- gated. Two algorithms for energy-based bearing estimation are developed. Their bias and variance expressions are derived. The study on bearing estimation shows that the perfor- mance (in terms of bias and variance) of the two energy-based bearing estimators depends on the omnidirectional signal-to-noise ratio (SNR), the spread in the gain profile of directional sensors and the angular separation between the look angles of two adjacent sensors in the cluster. ‘The range and depth estimators investigated are based on time difference of arrival measurements. The correlation among time difference of arrival (TDOA) estimators is in- vestigated. It is found that all the TDOAs, which are estimated with two time average auto-correlators and one time average cross-correlator, are correlated. The degree of corre- lation is expressed mathematically. ‘The effect of correlation among the multipath TDOA estimators on range and depth estimators is investigated using least squares range and depth estimators. Expressions for the variance and bias of the range and depth estimators are derived. It is shown that the theoretically predicted variance of both the range and depth estimators can be in error by a factor as much as 2.5 if the correlation among the TDOA estimators is not taken into account. All the theoretically derived results are corroborated with Monte Carlo simulations. BIOGRAPHICAL Born in Shanghai, China August, 1946 B.E. Electrical Engineering, Tsinghua University (Beijing) April, 1970 MSc. Electrical Engineering, University of Saskatchewan April, 1989, HONOURS University of Saskatchewan Graduate Scholarship April 89 - Aug. 92 PUBLICATIONS Y. X. Yuan and J. E. Salt, “Range and Depth Estimation Using a Vertical Array in a Correlated Multipath Environment”, IEEE Journal on Oceanic Engineering, Vol. 18, No. 4, pp 500 - 507, Oct. 1993. J. E. Salt and Y. X. Yuan, “On the Accuracy of Approximations of the Error Analysis of Registration Systems”, IEEE Trans. Aerospace and Electric Systems, Vol. 21, No. 6, pp 922 - 924, Nov. 1991 Y. X. Yuan, Analysis of ECG Body Surface Potentials, M.Sc. Thesis, University of Saskatchewan, Canada. iti Copyright In presenting this thesis in partial fulfilment of the requirements for a Postgrad- uate degree from the University of Saskatchewan, I agree that the Libraries of this, University may make it freely available for inspection. I further agree that permission for copying of this thesis in any manner, in whole or in part, for scholarly purposes may be granted by the professor or professors who supervised my thesis work or, in their absence. by the Head of the Department or Dean of the College in which my thesis work was done. It is understood that any copying or publication or use of this thesis or parts thereof for financial gain shall not be allowed without my written permission. It is also understood that due recognition shall be given to me and to the University of Saskatchewan in any scholarly use which may be made of any material in my thesis. Requests for permission to copy or to make other use of the material in this, thesis in whole or in part should be addressed to: Head of the Department of Electrical Engineering, University of Saskatchewan, Saskatoon, Canada S7N 0WO Acknowledgement The author wishes to express her sincere gratitude to Dr. J. E. Salt and Dr. G. C. Carter for their advice and guidance. Their assistance in the preparation of this thesis is especially appreciated. ‘The author would like to extend her appreciation and thanks to Dr. Daku, Mr. Macphedran and Mr. Jeffrey. Dr. Daku’s help with Monte Carlo simulations, Mr. PI Macphedran’s and Mr. Jeffrey's help with computer facilities were invaluable. Financial assistance was provided by Natural Science and Engineering Research Council and the University of Saskatchewan. This is gratefully acknowledged. The author would like to thank her parents, Bing Dong Yuan and Pei Fang ‘Yang, for their support. She also thanks her husband, Si Ming Pan, for his help and support during the study and her two lovely children, Li Yang Pan and Li Yuan Pan. for understanding the many years of her busy study schedule. UNIVERSITY OF SASKATCHEWAN Electrical Engineering Abstract 944400 PASSIVE LOCALIZATION OF AN UNDERWATER ACOUSTIC SOURCE USING DIRECTIONAL SENSORS Student: Y. X. Yuan Supervisors: Dr. J. E. Salt and Dr. G. C. Carter Ph.D. Thesis Presented to the College of Graduate Studies and Research October 1994 Abstract This thesis studies the passive localization (bearing estimation and range and depth estimation) of a single stationary underwater acoustic source in a two-path (direct and surface-reflected path) environment. The main objective of the study is to synthesize and then analyze a passive localization system with a two-receiver (vertically separated) structure that can be supported by a single sonobuoy. Each of the two receivers consists of a cluster of directional sensors uniformly located on a small circle in a horizontal plane. The gain profile of the directional sensors is such that it depends only on bearing of source. This novel receiver is proposed to facilitate bearing estimation. Two aspects of passive localization. bearing estimation and range and depth estimation. are studied separately with two different approaches. The source bearing is estimated from the average signal powers received at. two adjacent sensors in one of the two clusters. The focus in the study of bearing estimation is on the feasibility of energy-based bearing estimators. The range and depth information of the source is extracted from the 6 estimated time differences of arrivals (TDOAs) at the two receivers (which, for the purpose of range and depth estimation, are treated collec- tively as two omnidirectional sensors). The emphasis in the study of range and depth estimation is on investigating the correlation among the multipath TDOA estimators and its effect on the prediction of the variance of the time delay based range and depth estimators. ‘Two algorithms for energy-based bearing estimation are developed. Expressions for the bias and variance of both energy-based bearing estimators are derived. These expressions reveal a fairly simple relationship between the performance jin terms of bias and variance) of the two energy-based bearing estimators and the design parameters, i.e. the omnidirectional signal-to-noise ratio (SNR) and the two receiver parameters (the spread in the gain profile of directional sensors and the angular separation between the look angles of two adjacent sensors in the cluster). The results in the study of energy-based bearing estimation show that using the developed expressions can provide good predictions of the bias and variance of the two energy- based bearing estimators in the case of reasonable signal-to-noise ratios. ‘The range and depth estimators are investigated with time delay techniques. This study of time delay estimation investigates the correlation among the time differ- ences of arrivals (TDOAs) which are estimated with two time average auto-correlators and one time average cross-correlator. Fifteen expressions for the covariance among the six multipath TDOA estimators are derived. It is shown using the fifteen covari- ance formulas that all six multipath TDOA estimators are correlated and the degree of the correlation among the six TDOA estimators depends on the power spectral density of the source signal and that of the ocean noise, and only two multipath TDOAs, specifically the time difference of arrival between the direct and surface- reflected path for each receiver. The fifteen pairs of the six multipath TDOAs can be divided into three categories and a generalized covariance equation is developed for each category. Thus, the fifteen covariance expressions can be obtained from three compact expressions. ‘The effect of correlation among the multipath TDOA estimators on range and depth estimators is investigated in this thesis. The least squares range and depth estimators. based on all six TDOA estimators, are proposed. Expressions for the variance and bias of the range and depth estimators are derived. The results in the study of range and depth estimation show that the expressions for the variance of the range and depth estimators can be in error by as much as a factor of about if the correlation among the TDOA estimators is not taken into account. Including the effect: of correlation among the TDOA estimators provides significant'y better predicted range and depth variance. The results from this portion of study also show that the range and depth estimators are nearly unbiased. The expressions for the bias and variance of the two energy-based bearing es- timators, the expressions for the covariance of the multipath TDOA estima:ors. and the expressions for the variance and bias of the range and depth estimators, which incorporate the correlation in the TDOA estimators, are all corroborated with Monte Carlo simulations. Table of Contents Copyright iii Acknowledgements iv Abstract v Table of Contents vii List of Figures x 1 Introduction 1 1.1 Background of Passive Localization ...........-..-.-- 1 1.2 The Problem of Interest ©... 0.0.0 e eee eee eee 8 2 Problem Formulation 4 21 The Transmission System Model... 2.2.20 ee eee uw 21.1 Source Model ......- 5. 15 21.2 Ocean Acoustic Channel . . 15 2.1.3 Passive Localization System . . 2.14 A Composite Transmission Model . Oa Comey 2.3. Bearing Estimators . . 23.1 Algorithm One 290 Aotun Two) Seana so. 29 24 TDOA Estimators... 0. ee eee 30 2.5 Range and Depth Estimators 35 2.6 Model Summary 39 3 Performance of the Bearing Estimators 41 3.1. Bias and Variance Expressions... .. 2... .-.-00-0.0.-. 41 vil 3.1.1 Bias and Variance Expressions for Algorithm One ....... 41 3.1.2 Bias and Variance Expressions for Algorithm Two... .. - 46 3.2. Optimal Receiver Structure . . pees gee ae 48 51 3.3. Example Bias and Variance 3.3.1 An Example for Algorithm One . . . . 53 3.3.2 An Example for Algorithm Two... . . ae 54 3.4 Numerical Simulation 37 34.1 Simulation Configuration...........- 37 3.4.2 Comparison of Simulat 63 3.5 Summary ee 2 Theoretical Investigation of Correlation among the Multipath Time Difference of Arrival Estimators 74 4.1 Background of TDOA Estimation cece e cece ee TH 4.2 Covariance Expressions for the Multipath TDOA Estimators ..... 76 4.3 Correlation Coefficients of the Multipath TDOA Estimators ..... 82 44 Numerical Simulation for Correlation Coefficients . . 88 44.1 Simulation Configuration 38 44.2 Comparison of Simulation Results with the Theory ...... 94 4.5 Summary 7 Performance of the Range and Depth Estimators 99 5.1 Expressions for the Variance and Bias of the Range and Depth Estimators 99 Effect of Correlation among the TDOA Estimators 101 5.3 Numerical Simulation... .. is 5.3.1 Simulation Configuration...........- 108 5.3.2. Comparison of Simulation Results with the Theory ...... ILL 64 Siem lie Summary and Conclusions 120 References 128 Derivation of the Bias and Variance of the Bearing Estimators 131 A.L Bias and Variance for Algorithm One... . . « 131 A. and Variance for Algorithm Two . 138 Derivation of the Moments of ex, £141, 5; fy and & 142 B.1 The Expressions for the Second Order Moments of uey1, us, Yee and vy 142 B.2 The Moments of ¢x and eer... 2. - + 148. B.3. The Mean Value and Variance of S........-. fee 152 B4 The Moments of Zand... 2222022 ee ee eee a ky Derivation of the Covariance of the Multipath TDOA Estimators 162 C.1 Expression for the Covariance of D,,g, and Ds,4,-. - - - 1g) C.2_ Expressions for the Covariance of Dimim and each of Daydz, Darzzs Dns ie Deas . 164 C.3 Expressions for the Covariance among Day. Daisy, Daya; and Days, - 170 ‘The Derivations of the Partial derivatives 24, =¥%- and =3%= 176 Lat 12 2.3 24 25 2.6 27 28 29 2.13 214 215 217 3.1 List of Figures A Focussed Localization Beamformer. . . - ‘A Sonobuoy Type Passive Localization System. Model of the Source Signal. ... 22.2222 eee eens Modelling the Ocean Noise as NV Independent, identically Distributed Noises. Model of Directional Noises... . 2.2.2... 2-0 -050% A Receiver with Twelve Directional Sensor Elements Located on a Circle. Gain Profiles of Directional Sensor Elements in a Horizontal Plane. Gaussian Gain Profiles for Two Adjacent Sensors. ‘The Model for the Voltage Signal Delivered by a Single Directional Sensor. . . A Two-Receiver Vertical Array and an Acoustic Source in a Direct and Surface-Reflected Path Propagation Environment. The Gaussian-Shaped Gain Profiles for Two Adjacent Sensors in a Receiver as a Function of Normalized Horizontal Angle of Arrival. . . Model of the Bearing Estimator for Algorithm One. . Model of the Bearing Estimator for Algorithm Two.......-.-- An Illustration of Shifted Peaks in the Sum (solid line) due to the Effect of Superpositions for Two sinc Functions. . . . . Model of the Auto-Correlation TDOA Estimator. . - Model of the Cross-Correlation TDOA Estimator. Bis pestle Dated by Sis Tee 1D OAs fee the oes De 500 Meters and Range 500 Meters. Six Hyperbolas Defined by Six Noisy TDOAs for the Source at Depth 500 Meters and Range 500 Meters. ..........--.5 Model of the Range and Depth Estimators........... Theoretical Biases of Two Bearing Estimators versus the Fraction =~? when 8, = aies~ The Solid Lineis fr Algorithm One, while the Dashed Line is for Algorithm Two. . .. . oe 32 3.3 34 3.6 3.9 3.10 3.1 3.12 3.13 Theoretical Variances of Two Bearing Estimators versus the Fraction sata When 6, = pes. The Solid Line is for Algorithm One, while the Dashed Line is for Algorithm Two. . . . . : Theoretical Bias of the Bearing Estimator for Algorithm One. The Lines Indicate the Results Calculated from (3.6), while the Mark “+” Represents the Bias Obtained sing Simple Equation 3.8 for All Four Ranges. : Theoretical Variance of the Bearing Estimator for Algorithm One. The Lines Indicate the Results Calculated from (3.2), while the Marks “o and “+” Represent the Approximate Variances Given by (3.7) and (3.9) Respectively for All Four Ranges... ..-..--.---- Theoretical Bias of the Bearing Estimator for Algorithm Two. The Lines Indicate the Results Calculated from (3.15), while the Mark “+” Represents the Bias Obtained using Simplified Equation 3.17 for All Four Ranges. oe Theoretical Variance of the Bearing Estimator for Algorithm Two. The Lines Indicate the Results Calculated from (3.11), while The Marks “o” and “+” Represent the Approximate Variances Given by (3.16) and (3.18) Respectively for All Four Ranges. . . Monte Carlo Simulation Algorithm for the Bearing Estimators. Implementation of the Voltage Signal Generator ‘or the Bearing Esti- Matos ee Implementation of the Linear ee for Delayed Versions of the Source Signal. . . . The Additive Noise Generator for the Bearing Estimators. ... . . Biases for Algorithm One for jiss1 — He = @ = $ Radians, The Sim- ulation Results are Represented by the Marks, while the Theoretical Results from (3.6) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges... . . . Variances for Algorithm One for u41— ze = 7 = ¢ Radians. The Sim- ulation Results are Represented by the Marks, while the Theoretical Results from (3.2) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges..........---.- Biases for Algorithm One for fis: — 4x = 7 = 7 Radians. The Sim- ulation Results are Represented by the Marks, while the Theoretical Results from (3.6) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges... . . xi 33 Bt 66 3.14 3.15 3.16 3.7 3.18 4 42 43 4 47 43 49 Variances for Algorithm One for pst — x = ¢ = % Radians. The Simulation Results are Represented by the Marks, while the Theoret- ical Results from (3.2) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges..........- Biases for Algorithm Two for ues: — se = Radians. The Sim- ulation Results are Represented by the Marks, while the Theoretical Results from (3.15) are Presented as Solid, Dashed, Dotted and Dash- dot Curves for the Same Ranges. ....--..- +--+ Variances for Algorithm Two for fxs: — fk = Radians. The Simulation Results are Represented by the Marks, while the Theoret- ical Results from (3.11) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges........ : Biases for Algorithm Two for es: — ue= 9 = #5 radians. The Sim- ulation Results are Represented by the Marks, while the Theoretical Results from (3.15) are Presented as Solid, Dashed, Dotted and Dash- dot Curves for the Same Ranges. : Variances for Algorithm Two for ksi — He = 0 Radians. The Simulation Results are Represented by the Marks, while the Theoret ical Results from (3.11) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges... .. | : Theoretically Calculated Coefficient 75, 5,4: Theoretically Calculated Coefficients 95,4 5 and 2p, 4, ba,» tepresented by solid, dashed, d respectively. 2 PBaya, Doves id dashdot curves Theoretically Calculated Coefficients pp,., ee and 7,.,54,4« indicated by solid, dashod, Gehted ad dathdor Caetes respectively. ee Theoretically Calculated Coefficients 5,4, .,44: PBnegbayoq' PDayexDare Pi aasOn04' Paya Baseg 294 Pd Baeq! AC indicated by solid (bottom), dashed (bottom), solid, dashed, dotted and dashdot curves respectively. Monte Carlo Simulation Algorithm for Correlation Coefficients. Implementation of the Reosived Signal Generator for the Correlation Coefficient Estimation. .. . . . : ‘The Additive Noise Generators for the Correlation Coefficient Estima- Acne Implementation of the Auto-Correlation TDOA Estimator. . . Implementation of the Cross-Correlation TDOA Estimator. . 68 69 a7 39 90 a 92 93 4.10 Comparison of Simulation-Based Correlation Coefficient p,, Indicated by “+”, with Theory. .... 2... eos Coefficients 5, 4, busy: Indicated by “+”, “+”, Yo" and 4.11 Comparison of Simulation-Based Correl: Pong Dan’ Pre, Pay 4 P6,,4, Ba, espectively, with Theory... oo. ee ee oe 4.12 Comparison of Simulation-Based Correlation Coefficients 5,444, pen ge Pa Boba Indes by “4, 6", dd on sts withthe 7 4.13 Comparison of Simulation-Based Correlation Coefficients Ap,,,,,, Pongdan? Pundan’ ag bns? Page Pag! 24 Pu gdyog’ Hl cated by + (boutom), “+ tottom), #2,", Yor anda Re tively, with Theory. . 98 5.1 Theoretically Obtained Range Variances for = = 200 Meters and 2 = 400 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . 102 5.2. Theoretically Obtained Depth Variances for =, = 200 Meters and 2 = 400 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . . . 103, 5.3. The Ratios of the Range Variances (solid line) and the Depth Variances (dashed line) without and with Considering the Correlation among the ‘TDOA Estimators for z; = 200 Meters and =: = 400 Meters. = 104 5.4 Theoretically Obtained Range Variances for z; = 100 Meters and 300 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . 5.5 Theoretically Obtained Depth Variances for =, = 100 Meters and 1 300 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . 105 5.6 The Ratios of the Range Variances (solid line) and the Depth \ Variances (dashed line) without and with Considering the Correlation among the ‘TDOA Estimators for z, = 100 Meters and =z = 300 Meters... . 106 ‘Theoretically Obtained Range Variances for =: = 200 Meters and =: = 900 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . . : 106 5.8 Theoretically Obtained Depth Variances for + = 200 Meters and =1 = 600 Meters: the Dashed Line Assumes Independent TDOAs, while the Solid Line Includes the Effect of the Correlation. . . : 107 5.9 The Error Ratios of the Range Variances (solid line) and the Depth Variances (dashed line) without and with Considering the Correlation among the TDOA Estimators for z: = 200 Meters and 22 = 600 Meters. 5.10 Monte Carlo Simulation Algorithm for Range and Depth Estimators. 5.11 Comparison of the Theoretical Variance (solid line) and the Simulation- Based Variance (dashed line) of the Range Estimator for =, = 200 Meters and z, = 400 Meters. . a 5.12 Comparison of the Theoretical Variance (solid line) and the Simulation- Based Variance (dashed line) of the Depth Estimator for 2, 0 Meters and zz = 400 Meters. 5.13 Comparison of the Theoretical Bias (solid line) and the Simulation- Based Blas (dashed line) of the Range Estimator for 1 = 200 Meter and 22 = 400 Meters. - ce 5.14 Comparison of the Theor Based Bias (dashed line) of the Depth Estimator for =; and =, = 400 Meters. ee eae Comparison of the Theoretical Variance (solid line) and the Simulation- Based Variance (dashed line) of the Range Estimator for z1 = 100 Meters and 2. = 300 Meters. 5 : 5.16 Comparison of the Theoretical Variance (solid line) and the Simulation- Based Variance (dashed line) of the Depth Estimator for 1 = 100 Meters and z2 = 300 Meters... . : al Bias (solid line) a and the Simulation- 200 Meters Comparison of the Theoretical Bias ‘old line) and the Simulation- Based Bias (dashed line) of the Range Estimator for =, = 100 Meters and 2; = 300 Meters... 2. 5.18 Comparison of the Theoretical Bias (solid line) and the Simulation- Based Bias (dashed line) of the Depth Estimator for =: = 100 Meters and 2: = 300 Meters... . . : 5.19 Comparison of the Theoretical Variance (solid li \) a and the Simulation- Based Variance (dashed line) of the Range Estimator for 21 = 200 Meters and =, = 600 Meters... .... . : 5.20 Comparison of the Theoretical Variance (solid line) and the Simulation- Based Variance (dashed line) of the Depth Estimator for =, = 200 feters and = = 600 Meters. . . 5.21 Comparison of the Theoretical Bias (solid line) and the Simulation- Based Bias (dashed line) of the Range Estimator for 2 = 200 meters and 2, = 600 meters. . . 5.22 Comparison of the Theoretical Bias (solid line) and the Simulation- Based Bias (dashed line) of the Depth Estimator for =) = 200 Meters and =; = 600 Meters. . os 107, 109 113 ud 14 116 16 us 1. Introduction 1.1 Background of Passive Localization ‘The passive localization of an underwater acoustic source has received con- siderable attention in the literature. In a passive sonar system, the location of the underwater acoustic source is estimated from the “noise-like” acoustic signal radiating from the source. Location estimates obtained using this information are inherently prone to error. Knowing the accuracy of the localization system is essential. especially in military applications where actions taken require accuracy. Passive localization is not restricted to the field of underwater acoustics, it also has numerous applications in navigation, surveillance and geophysics. Techniques used to passively locate underwater acoustic sources are continu- ously changing. This is due in part to the changes in the radiation power of acoustic sources of interest, For example, a submarine using a nuclear power plant and modern technology to minimize vibration radiates a much different signal than the submarines used in World War II. As modern submarines become quieter, scientists are challenged to develop new more sophisticated localization algorithms to provide accurate source location estimates in environments where signal-to-noise ratios are ever decreasing tu. Passive localization of an underwater acoustic source often starts with detect- ing the presence of a distant source. This coarse localization process can be done with two well separated ships each towing a long line sensor array. The length of the line array could be more than 500 meters and the line array could contain hundreds of omnidirectional sensors. Each of the two towed long line arrays is electronically steered to get a highly directional gain profile. The steering referred to as beamform- ing is accomplished by delaying and summing the signals received from each sensor in the towed line array. The beamformer is swept to find the angle of maximum signal strength which gives the bearing of the source. The intersection of the two bearings from the two ships indicates the location of the source. When the source is several kilometres from each ship, errors in the bearing estimates result in a coarse estimate of the source location. The details of this type of localization have been studied extensively (1, 2, 3, 4, 5} There is a method of beamforming that not only estimates bearing but also es- timates range. With this type of beamforming, referred to as focussed beamforming, only one towed line array is required to get a coarse location estimate (bearing and range). Figure 1.1 shows the focussed beamformer proposed in [6]. The sensor signals entering the beamformer are filtered in a Wiener-like manner, appropriately delayed according to a hypothesized source position (bearing and range), combined and then the output power is calculated. The hypothesized source position that yields the max- imum output power is taken to be the source position estimate. This gives estimates of both bearing and range, however, the range estimate is usually very inaccurate for distant sources [2]. There is a variety of configurations of focussed beamformers depending on range and effective Laseline [6]. Unfortunately, all are computationally intense as they have a two dimensional search space. 2| Prefilters Delay L tt : -| Prefiters =z Delay Beamformer oa 2 zs Squares and |_Output . A Average e | Prefileers Delay M M Hypothesized |< Position Delay Computation Figure 1.1: A Focussed Localization Beamformer. Alter the source is detected and coarsely localized in a certain region of the ocean, then a more accurate location estimate can be obtained using sonobuo; a sensor arrays, which are deployed in the vicinity of the detected location. A sonobuoy is a floa: mechanism which supports a vertical line array. The length of the array is at most a few hundred meters. The sonar signals picked up by the sensors in the array are connected to the mother ship via a radio frequency transmitter. The line array, transmitter, and cables can be packaged in a cylinder that can be easily transported and deployed. These sonobuoy-type sensor arrays can be easily dropped in a certain region of the sea using a helicopter. They are not normally retrieved for re-use so are usually very simple and contain relatively few sensors. A block diagram of a sonobuoy-type localization system is shown in Figure 1.2. Prefilter Correlator L = ——+[ Prefitter. |_4—=f Corelator eee 7 ition Parameter Sensor 2 Estimates Signals e e Boats e e e Prefilter Correlator - eit A Sonobuoy Type Passive Localization System. Figure ‘The signals received at submerged sensors are filtered, the time differences of arrivals (TDOAS) of the signals at the sensors are measured with pairwise correlators, and the source location information is then extracted from these estimated time differences of arrivals (TDOAs) (7, 8, 9]. This structure is preferred in sonobuoy scenarios where there are relatively few sensors in the array. When the sensors are submerged, there are several paths that link the source to the sensors due to the presence of the boundaries or inhomogeneities in the real sea medium. For example, the sensors may receive the underwater acoustic energy through direct paths, surface-reflected paths, bottom-bounce paths, and all their combinations. In a scenario where the source is quite far from the sensors, most, of the information is contained in the direct paths. The early work on localization problems concentrated on this scenario (10, 11, 12, 8, 2, 6, 9, 5]. In conventional localization where omnidirectional sensors are used in a receiver array, the geometric information of the source location is fully encoded in the time differences of arrivals (TDOAs) at the spatially separated sensors. Three or more sensors, at least two of which must be horizontally separated, are required to accomplish three dimensional localization. The time differences of the signals arriving at a pair of sensors can be estimated with various time deley estimation procedures. Collected reprints and a bibliography of work in this area are available in [7]. One of these procedures is implemented by cross-correlating the two received signals. For an M-sensor system in a direct-path- only propagation environment, cross-correlators can provide a set of M(M — 1)/2 distinct time difference of arrival (TDOA) estimates. A constant sound speed ocean medium is generally assumed in the literature, which means the time delays between received signals at two or more sensors are proportional to the distances between receiving sensors and the source of interest. Each of the time differences of arrivals (TDOAs) defines a hyperbola (or a hyperboloid in a three dimensional space). The source location can be found from the intersections of the set of hyperbolas (or hy- perboloids). Finding the hyperbolas and then the intersection point is a complicated non-linear operation. In general, the computation of the location estimates using 2 multisensor array can be quite cumbersome and expensive. ‘The assumption that the acoustic energy propagates from the source to the receiver array along the direct-path-only simplifies the localization problem. But this direct-path-only propagation model does not reflect reality. A more complicated multipath propagation model better resembles the real ocean environment. Recently, the effect of the multipath self-interference on passive localization has caught the interest of scientists. Evaluation of the performance of localization systems based on multipath information has gained attention in the recent literature. Evaluating the performance of locating an underwater acoustic source in a mul- tipath environment has been the subject of many studies (13, L4, 15, 16, 17, 18, 19. 2 21). Abel and Smith (17] have studied range and depth estimation using a single sensor in a M-path environment. This study has developed range and depth estima- tors as the minimizers of weighted equation errors. These estimators are closed-form functions of time difference of arrival (TDOA) estimators and therefore are more com- putationally efficient than that offered by iterative type estimators. The accuracy of their range and depth estimators was determined under the assumption that the time difference of arrival (TDOA) estimators are independent, zero-mean, Gaussian ran- dom variables. Research done by Rendas and Moura [15] has investigated the accu- racy of localization under very general conditions. In their work, a general expression for the Cramer-Rao lower bound on the localization error was derived for multiple sources in 2 M-path environment. It was concluded in (15] that the contribution of the multipath delays can be understood as the results of the spatial processing of a virtual array. About the same time, results of a study focussing on passive ranging in a two-path environment were published from Hamilton and Schultheiss’ work (14]. As done in [15], the accuracy of range estimation was analyzed using the Cramer-Rao lower bound on multipath time difference of arrival estimators. It was also pointed out in [14] that the passive range estimators in multipath dominant environments offer significant performance improvement over conventional methods which assume a direct-path-only propagation environment. Utilizing the multipath information is certainly beneficial in predicting the location of the source. It is noted that conventional localization estimators are time delay based. There- fore, the statistical uncertainty in localization estimators is closely related to the accuracy of time delay estimators. The subject of time delay estimation has been well explored (7, 2]. Among these studies, relatively few works have addressed the time delay estimation in a multipath environment compared to the direct-path-only environment. ‘The Cramer-Rao lower bound covariance matrix of the multipath time difference of arrival estimators (CRLB(D)) has been developed in [14, 15, 19]. It has been concluded in their studies that the CRLB(D) is generally not diagonal. However, in much of the literature on the topic of performance evaluation of multipath localization, this finding has been ignored and diagonal covariance matrices have been assumed (13, 17, 20]. Perhaps the reason for this is that the diagonal elements of the covariance matrix, that is, the variances of errors in the auto- and cross-correlation multipath time difference of arrival (TDOA) estimators, have been available in the literature (22, 23, 13]. In summary, two approaches are taken in the performance evaluation of local- ization systems. One is to use the Cramer-Rao lower bound covariance matrix of the multipath time difference of arrival estimators (CRLB(D)) such as was done in (14, 15, 19]. In this approach, the correlation among the time difference of arrival (TDOA) estimators is taken into account. The other is to use a diagonal covariance matrix of the multipath time difference of arrival (TDOA) estimators where the di- agonal elements are the variances of the time difference of arrival (TDOA) estimators [13, 17, 20]. In the later case, while the diagonal variances may be more accurate than those in the CRLB(D), the off diagonal elements, which could be non-zero, are ignored. Thus, the covariance matrix approach ignores the correlation among the time difference of arrival (TDOA) estimators. The question that is left unanswered in the literature is whether or not the correlation among the time difference of arrival (TDOA) estimators can be neglected. In addition to the assumption that the sound speed profile is constant, most work in the literature on conventional localization also assumes: the ambient ocean noise is non-directional [24], the attenuation of the signals is dominated by spherical spreading, the surface and bottom reflections are specular, and the ocean medium is homogeneous. These assumptions are reasonable when the underwater acoustic source of interest is in the near vicinity of the receiver array, say within 10 km (25. 26]. It is generally understood that the performance of the localization system is often degraded in a real sea environment. For example, the sound profile. the most. significant parameter for localization estimation, is a complicated function of temper- ]. The ature, time, depth and the physical and chemical properties of the sea (25 noise from marine organisms, sounds of the wave action, storms, man-made sources. and many other noise sources are directional in nature. Spherical spreading may not dominate when the source is on the surface or bottom [25]. Sound energy is also absorbed due to thermal conductivity, viscosity, structural and chemical relaxations etc (26, 28]. If the ocean surface is not perfectly smooth, the incident sound waves do not experience pure specular reflection and have some scattering. 1.2 The Problem of Interest The problem of interest in this study is the passive localization (estimating bearing, range and depth) of a single stationary underwater acoustic source in a two-path (direct and surface-reftected) ocean environment. The localization system consists of a vertical array supported by a single sonobuoy. The study assumes that the source is near the receiver array and that the ocean is deep. When the ocean is deep, the bottom-bounce acoustic energy is very small and therefore ignored. ‘Most of the work to date has concentrated on two dimensional localization either estimating range and bearing or estimating range and depth. This is mainly because the passive localization of a source in a two dimensional space is relatively simple. In the case of range and depth estimation, the bearing of the source is assumed to be known. In this situation, the range and depth estimation can be accomplished using a single sonobuoy which supports a vertical line array with two or more omnidirectional sensors. However, if the source bearing is unknown, at least two horizontally separated sonobuoys are required to acquire bearing information. This thesis proposes a two-receiver vertical array to achieve 3 location parame- ters (bearing, range and depth). Each of the two receivers in the vertical array consists of a cluster of directional sensors which have their gain profiles highly dependent on angle of arrival. This means that the bearing information is contained in the relative energies of the signals received at directional sensors in one of the two clusters. In other words, it may be possible to realize energy-based bearing estimation using one of the two vertically separated receivers. Then, the localization system with both functions of bearing estimation and range and depth estimation can be implemented using a two-receiver vertical array which can be supported by a single sonobuoy. ‘The focus in the study of bearing estimation is to determine if the bearing of the source can be estimated from the signal energies received at directional sensors in one of the two clusters. The intent of the energy-based bearing estimation study is to 10 suggest an architecture for a vertical array that can be deployed as a single sonobuoy system and used to localize an underwater acoustic source in a three-dimensional space. Whether or not the particular gain profile that is used for the directional sensors in this thesis is realizable is not addressed. Using signal energy information instead of time delay information for bearing estimation is a novel approach in the field of underwater source localization. There is a precedence for this energy-based approach in the area of radar where bearing estimation was attempzed using average signal powers received at directional antennas [29]. The situation here differs from the radar system in that the acoustic source signal is propagating along a multipath channel and that the noises received at the directional sensors in one of the two clusters may be correlated. However, the energy- based bearing estimation has the possibly great advantage of utilizing the tonal power generated from the acoustic source. The feasibility study of bearing estimation using directional sensors in this thesis includes the development of energy-based bearing estimators in an underwater environment and the investigation of the relationship between the accuracy of the energy-based bearing estimators and the structure of the cluster of directional sensors. ‘The accuracy of energy-based bearing estimators is measured by their variance. The bias is also important but if the bias of an estimator can be predicted, then it can be removed and is not an issue in the accuracy of the estimator. However, the uncertainty which is measured by the variance cannot be eliminated. The feasibility of novel bearing estimators using an energy-based approach is established by showing that the bias of the energy-based bearing estimators can be predicted and therefore i removed and that their variance is small. It is desirable to express both the bias and variance of the energy-based bearing estimators in terms of the design param- eters that include the omnidirectional signal-to-noise ratio (SNR) and the receiver structure parameters. Therefore, the number and type of directional sensors required in a cluster can be determined. This is a complicated task if the noises received at directional sensors in one of the two clusters are correlated. The range and depth estimation of interest is investigated in a conventional way using multipath time delay techniques. For the purpose of time delay based range and depth estimation, the two clusters of directionai sensors in the proposed vertical array can be collectively treated as two omnidirectional sensors. The time delay-based range and depth estimation in a two-path environment using a two-element vertical array has been investigated in [19], where the accuracy of range and depth estimators was measured using the Cramer-Rao lower bound on the multipath time difference of artival estimators (CRLB(D)). The Cramer-Rao lower bound covariance matrix so obtained is not diagonal. This is the same conclusion obtained in (14, 15] which is reviewed in the previous section. It is noticed that the Cramer-Rao lower bound of a multipath time difference of arrival estimated with a time average auto-correlator in a two-path environment is very optimistic when the surface-reflected signal is received with nearly the same strength as the direct path (see (22, 23]). Therefore, as pointed out in [19], better estimates of range and depth should be obtained using the actual error covariance of the time difference of arrival estimators. It was introduced in the previous section that, for the performance evaluation of multipath localization, the covariance matrix of the multipath time difference of arrival estimators is usually assumed to be diagonal. The focus in the study of range and depth estimation in this thesis is on finding the correlation among the multipath time difference of arrival estimators and its effect on the range and depth estimation error. The primary reason for this focus is the absence of expressions for the covariance of multipath time difference of arrival estimators in the open literature. ‘The study of passive localization estimation, which includes 3 location param- eters (bearing, range and depth), using a two-receiver vertical array supported by a single sonobuoy is reported in this thesis. The remaining five chapters are briefly described below. Chapter 2 presents the mathematical models for the development of the theoretical bias and variance expressions for two energy-based bearing esti- mators and time delay-based range and depth estimators. Chapter 3 evaluates the performance of the two energy-based bearing estimators. Expressions for the bias and variance of the two energy-based bearing estimators are derived. The relation- ship between the performance of the two energy-based bearing estimators and the design parameters (i.e. the omnidirectional signal-to-noise ratio and the receiver structure parameters) are analyzed using the developed expressions. The theoreti- cal development of the two energy-based bearing estimators is finally corroborated with Monte Carlo simulations. Chapter 4 contains the theoretical investigation of the correlation among the multipath time differences of arrivals estimated from two time average auto-correlators and one time average cross-correlator in a direct and surface-refiected path environment. Expressions for the covariance of the multipath time difference of arrival estimators are derived in terms of either the power spec- tral densities of the received signals or that of the source signal and noise. explicitly. 13 This chapter closes with a Monte Carlo simulation that corroborates the theory. In Chapter 5, expressions for the variance and bias of the range and depth estimators are derived based on the covariance matrix of multipath time difference of arrival (TDOA) estimators that is determined using the covariance expressions developed in Chapter 4. The effect of correlation among the multipath time difference of arrival (TDOA) estimators on range and depth estimators is illustrated with numerical ex- amples. Again, the development of the theoretical variance and bias expressions is corroborated by Monte Carlo simulations. The thesis ends with the summary and conclusions of Chapter 6. 2. Problem Formulation ‘The localization of an underwater acoustic source studied in this thesis work is accomplished in two phases: the implementation of the energy-based bearing esti mators and the realization of the time delay based range end depth estimators for a given bearing in a two dimensional space. The acoustic source of interest in this study is near, say within 10 km, the two-receiver vertical array ina deep water environment, where it is reasonable to assume the presence of only the direct and surface-reflected paths This chapter establishes the mathematical models fo: the analysis of the accu- racy of the bearing, range and depth estimators. A model for the physical transmis- sion and reception of the acoustic signal in an underwater multipath environment is introduced in Section 2.1. Then, Section 2.2 presents the geometry for a localiza- tion systern employing a two-receiver vertical array. The subsequent three sections introduce the models for the energy-based bearing estimators, the time difference of arrival (TDOA) estimators, and the time delay based range and depth estimators, respectively. Finally, this chapter is concluded with a summary. 2.1 The Transmission System Model The transmission system for a passive localization problem in an underwater acoustics environment consists of three components: the source, the ocean acoustic ry Gaussian w(t) White Noise HAS) Generator s(t) Figure 2.1: Model of the Source Signal. channel and the localization system. The details of the physical description for these three components can be found in [27, 25]. This section describes the mathematical models of these three components for the problem of interest here. The source model and at least part of the ocean channel model are similar to those used in (16] and therefore are only briefly introduced. The localization system consists of two verti- cally separated directional receivers. The directional receiver is a novel concept and therefore is described in detail. 2.1.1 Source Model The underwater acoustic source of interest in this study is a submarine. It is assumed to be sufficiently far from the two-receiver vertical array to be modelled as a point source. It is also assumed to have an omnidirectional radiation pattern which is a stationary, Gaussian random process with a known broadband spectrum. The model of the source signal is shown in Figure 2.1. A white Gaussian noise generator produces the time signal w(t) which is filtered to produce the voltage signal s(t) emanating from the source. The frequency response of the filter is denoted H,(f). 2.1.2 Ocean Acoustic Channel There are four main factors which affect the acoustic signal while it is propa- gating from the source to the two receivers in the vertical array in a noisy oceanic environment. These four factors are the sound speed profile, the two propagation 16 paths, the sound loss during the propagation, and the ambient noise. Since the source is assumed to be reasonably close to the receivers, the complex nature of the physical phenomena can be simplified to facilitate the analysis of the localization system. The assumptions for the first three factors (which are also made in (16]) are briefly described below: 1. The sound speed is assumed to be constant with value c = 1,465 meters per second and so does not depend on the ocean depth or temperature nor does it, depend on the frequency of the acoustic signal; 2. Both direct and surface-reflected propagation paths exist and the reflection from ocean surface is assumed to be specular. This causes a phase reversal at the surface boundary; 3. The bottom-bounce paths are assumed to be highly attenuated; either the spreading loss nor the absorption loss are dependent on frequen 5. The spreading is spherical. For a uniform radiating point source, the spreading loss (power) is inversely proportional to the square of the distance from the source to the point of interest [25]; 6. The absorption, which is due to the conversion of acoustic energy into heat, is taken to have value a, = 0.2187 x 10-°dB/meter. With these assumptions, the intensity of the acoustic signal travelling over path i with length {; is proportional to a; where a = ehio-Beseeh, (2.1) Figure Modelling the Ocean Noise as N Independent, Identically Distributed Noises. There are two possible paths: the direct and surface-reflected paths. These are dis- tinguished through the subscript d or 5, i.e. ag or a. The sign in (2.1) is dependent on the path, being negative for the surface-reflected path and positive for the direct, path. The coefficient a; is referred to as the attenuation coefficient. Another factor is the ambient ocean noise, which is the root of the localization error and therefore is the most important factor in the localization problem. The ocean noise is usually modelled as an ensemble of identical noise sources uniformly distributed throughout a sphere with a large radius [24]. For the purpose of this study, the reasons for which will be explained later, the noise is modelled as NV (which is a very large number, approaching infinity) point noise sources located on a circle in the horizontal plane of the ocean surface. All of these noise sources are Gaussian, independent, identically distributed random processes with zero means and identical power spectra. Each noise source emits just one ray directed at the receiver as shown in Figure 2.2. The total noise at the point of interest is the sum of these independent, identically distributed noise sources. It is also assumed that each of the noise sources 18 Gaussian | White Noise to) Generator 0 An(f) G i w(t, Or White Note --2&4) Half) Generator 1 Gaussian White Noise Generator N — 1 (ts va) | Half) Figure 2.3: Model of Directional Neises. is independent of the source signal. ‘The complete model for the ocean noise is shown in Figure 2.3, where .V white noise generators produce N rays of independent, Gaussian, white noise sources w(t. ;) for i € {0.1.2....,.V — 1} with their angular separation of 8; = 41. Each of them is then filtered using identical filters, whose frequency response is denoted H,(w). to produce broadband noise sources “#1. The total noise impinging on a sensor. denoted n4(t), is expressed mathematically by 2.1.3. Passive Localization System A eceiver in a localization system, which is used to receive underwater acoustic sound, incorporates an array of passive sensors, consisting of the preamplifier, pream- plifier housing, auxiliary circuits, cables and cable connectors. The sensors convert the acoustic energy to electric energy using piezoelectric or magnetostrictive elements. 19 ve Figure 2.4: A Receiver with Twelve Directional Sensor Elements Located on a Circle. Usually, the conversion is linear and has no directional preference. Sensors with these characteristics are called omnidirectional sensors. Sensors that receive signals with directional preference are called directional sensors. In this study, the passive localization, which is the estimation of bearing, range and depth, is accomplished using a system with two vertically separated receivers. Each of the two receivers in the vertical array consists of M directional passive sensors uniformly located on a circle as shown in Figure 2.4. When properly deployed. the x and y axes shown in Figure 2.4 are in a horizontal plane beneath the surface of the ocean. The diameter of the circle is assumed to be very small compared to the source range. It is also assumed that the directional gain profiles of the sensors depend on angle of arrival (or the source bearing) and not on the elevation angle. It is for this reason that the noise model, established in the previous section, is simplified to point sources on a horizontal plane. ‘The directional gain profile of the sensors in the two clusters is taken to be an even function of the horizontal angle with respect to the look angle of the sensor. which is corresponding to the largest value of its gain profile. The gain profiles of the sensors are chosen to overlap each other so that at least two sensors can receive significant levels of the signal over the observation time. For the purpose of the bearing estimation study in this thesis, the shape of the voltage gain profile is taken to be Gaussian truncated at +r radians. The gain profile of sensor & (k € {1,2,...,.M}) is expressed as a function of the horizontal angle of arrival with respect to its look angle. For a source arrival angle 8, and a look angle x, the voltage gain of sensor & is defined by (02) = -r (ah) 222, (26) ors eee ‘The sensors in the receiver are designed to be suffciently close together for each point source noise ray to impinge on all sensors. This is the worst case scenario : Gaussian Gaussian Gaussian Gaussian || White Noise White Noise White Noise |@ @ @| White Noise ; Generator Generator 0 Generator 1 kGenerator = 1 ; HAS) HAf) |@°e) HA |: | Delay Delay ' a Dis : ag as ‘ Additive Noise a(t) Figure 2.7: The Model for the Voltage Signal Delivered by a Single Di- rectional Sensor. since it implies worst case correlation between the noise voltages delivered by the two adjacent sensors. The correlation coefficient of the noise voltages delivered by the two adjacent sensors is approximately given by Sen 99" )g(O**")dO* Poems Sat OE I, Oa For the Gaussian shaped gain profile assumed in this work, the correlation coefficient becomes (2.8) 2.2 Geometry ‘The localization system proposed in this thesis employs two vertically separated receivers, each of which consists of a cluster of directional sensors. This two-receiver vertical array can be supported by a single sonobuoy. The geometry of the two-receiver vertical array and source positioning is illustrated in Figure 2.8. The two receivers are at depths =: and zz (the deeper receiver being at 22) and are mounted so that the sensors in the two receivers have the same orientation. A Cartesian coordinate system with the origin at the ocean surface above the two receivers is used. The zy plane is the ocean surface, and the vertical axis through the centers of both receivers. which is the = axis, is positive down. The z - y chosen in an early figure would have += up for ar. h. c. system. The source position is given by the vector, [24 ys =]. and has bearing 0, given by 8, =tan-' £. If the bearing is known, localization is reduced to in a particular vertical a two-dimensional problem since it is known that the source i plane. In this case, the line labelled x in Figure 2.8 is used as the horizontal axis. ‘The source position vector for a known beating is denoted by vector ps =[\s 21] -2 Image of Receiver 2 Image of Receiver 1 2 Receiver 1 Receiver 2 Acoustic Source (2s ys za]? oF [xe Figure 28: A Two-Receiver Vertical Array and an Acoustic Source in a Direct and Surface-Reflected Path Propagation Environment. where x, is the horizontal distance of the source from the origin, i.e. range. and =, is the depth of the source. It is assumed in Section 2.1.2 that the ocean surface is specular and the sound speed is constant. Therefore, the surface-reflected paths can be viewed as the signals collected by the sensors in two virtual receivers (images of the two receivers) located above the ocean surface as shown in Figure 2.8. It is also pointed out that since the cluster of sensors in a receiver is located in a horizontal plane on a circle with a diam- eter that is much much smaller than the source range, the time delay of arrivals along the direct or surface-reflected path is the same for all sensors in the same receiver. For a two-receiver array in a direct and surface-reflected propagation environment. there are only four paths. These paths have delays Da,, Diz, Ds, and D,,, where Dz, and D,, for j being either 1 or 2 are the time delays for the signal travelling from the source to the sensors in receiver j along the direct path and surface-reflected path respectively. The voltage signal delivered by sensor & in receiver j in accordance with Figure 2.7 is given by z(t) = 9(08) (24,s(t — De,) + 04,5(¢- Da,)) + nil), (2.9) where s(t) is the source signal, n;,(t), given by (2.6), is the noise voltage delivered by sensor k in receiver j, a4, and a,, express the attenuations experienced by the source signal in travelling from the source to receiver j along the direct and surface-reflected paths respectively. Guinn a o 05 t 1s Normalized Source Bearing Figure 2.9: The Gaussian-Shaped Gain Profiles for Two Adjacent Sensors in a Receiver as a Function of Normalized Horizontal Angle of Arrival. 2.3. Bearing Estimators ‘The gain profile of the sensors in the two vertically separated receivers depends only on the horizontal angle of arrival (or the source bearing) and not on the vertical angle of arrival as assumed in Section 2.1.3. The Gaussian shaped gain profiles of two adjacent sensors of interest, for example, 9(8}) and (63) for sensor 1 and sensor 2 in Figure 2.5, are plotted versus the normalized horizontal angle of arrival (source bearing) and represented by solid and dashed curves respectively in Figure 2.9. The horizontal arrival angle is expressed as a fraction of the angular separation between the look angles of the two sensors, i.e. —%—, which is referred to normalized source beating. The Gaussian-shaped gain profiles shown in the plot have a spread factor (standard deviation parameter in a Gaussian density), 7, equal to the angle of the separation between the look angles of the two sensors. Only one of the two receivers in the vertical array is required to estimate the bearing. Only two sensors in the chosen receiver whose look angles are closest to the bearing of the source are used for bearing estimation. Thus for the chosen sensors. the source bearing in terms of normalized arrival angle is restricted to the interval (0.1] in Figure 2.9. Two bearing estimation algorithms are developed in the following two subsections. 2.3.1 Algorithm One In the absence of noise, the ratio of the average signal powers delivered by two adjacent sensors of interest is a function of the source arrival angle 8, and is given by n= He) 1 POR’ magi entgeay oa ,M}. A primary where. for the sensors of interest, wx < 8 < past for k € {1,2 reason for assuming a Gaussian shape for the gain profile is that it facilitates the inversion of (2.10), which can be expressed as 0, = HahMess ¢ ay An 2 Wea eu) The ratio of the average signal powers delivered by the two sensors can be estimated by measuring the power delivered by each sensor and then taking their ratio, i.e., Bea’ ze(t) Be ORat aenlt sel) dy zone Bet Figure 2.10: Model of the Bearing Estimator for Algorithm One. where pg is given by Lf he = zh ai(t)dt (2.13) with T being the observation interval. It is noted that the subscript j in (2.9) is omitted because the analysis of bearing estimation involves only one receiver. An estimate of the bearing of the source, denoted 6,, is obtained from (2.11) using fi, given by (2.12), for rr. The model of this energy-based bearing estimator is given in Figure 2.10. 2.3.2 Algorithm Two ‘The arrival angle can also be estimated using the ratio of the difference and sum of the average signal powers delivered by the two adjacent sensors, sensor k and sensor k + 1, whose look angles are closest to the arrival angle. This ratio estimator is given by ¥, —, 2.14) * Pat Best oe which in the absence of noise can be expressed as 2(08) — g2(et4 (08) = 9°(05**) ae ne 9°(85) + g2(85*")” 30 Os ryrae Inverse, Lt Function eei(t) Je (Pat Pret Figure 2.11: Model of the Bearing Estimator for Algorithm Two. For Gaussian shaped gain profiles, (2.15) can be inverted to give the bearing of the source with the result a, = Het Mee : =rs 216 2 + ima) oS ‘An estimate of the bearing of the source can be obtained from this expression using an estimate of the power ratio given by (2.14). The model of the energy-based bearing estimator here is illustrated in Figure 2.11. 2.4 TDOA Estimators ‘The range and depth of the source are estimated from the differences in arrival times at the two vertically separated receivers. Only one sensor from each receiver is used in the estimation of the time differences of arrivals (TDOAs). The sensor used in each receiver is the one with its look angle closest to the source bearing. ‘As indicated in the previous chapter, much work has been done on TDOA estimation and is summarized in the recent text by Carter [7]. The Cramer-Rao lower bound matrix of the multipath TDOA estimators for a two-sensor vertical array 31 in a two-path environment has been developed by Friedlander [6]. The off-diagonal elements of this CRLB matrix are non-zero, which suggests the TDOA estimators may be correlated. However, most work on the performance evaluation of localization estimators in a multipath environment has assumed that the errors in the TDOAs estimated from auto- and cross-correlators are statistically independent of each other [13, 17, 20]. The validity of this assumption is examined in this thesis. The focus of the TDOA estimation study in this thesis is on the correlation among the time difference of arrival (TDOA) estimators and its effect on the range and depth estimation. The effect of the gain profile of the sensors on the TDOA estimation is not a big issue and therefore not addressed in this study. This allows each of the two receivers in the vertical array to be modelled, for the purpose of the TDOA esti ation analysis only, as a single omnidirectional sensor. For a two-receiver vertical array in a direct and surface-reflected propagation environment, the geometric information of the source position is encoded in the dif- ferences among the path delays Ds, Da, Ds, and D,. The four paths yield six time differences of arrivals (TDOAs): Dy, — Day, Daz ~ Daz, Day ~ Dazy Dry — Das Dz, — Daz, Dé, — Dag. To simplify notation, these TDOAs will be denoted by a double subscripted D, for example D,,4, for difference D,, ~ Da. The six TDOAs of interest are parameters in two auto- and one cross-correlation function. The autocorrelation functions of the voltage signals produced by the first and second receivers contain parameters Dy, and D,,4, and are given by Rrnm(T) E{zm(t}zm(t—7)} (03, +02) Ral) + Gag sm (Rat + Dams) + Ralt = Dimam)) + Bolt), (2.17) where m is either 1 or 2 for the signal from receiver 1 or 2 respectively, E{«} denotes the expected value, R,(r) = E{s(t)a(t—r)} and Ro(r) = E {no(t)no(t — 7)} are the statistical auto-correlation functions for the source signal s(t) and noises ng(t) (which is given by (2.2), and as previously explained, Duns, is the difference Ds, ~ Dim: The cross-correlation function of the voltage signals produced by the two re- ceivers contains parameters Dayés, Dain Dads and De,»y, and is given by Ralt) = E{zi(t)za(t-7)}, = 04,04, Ri (7 — Deydz) + On On R(T — Dayz) + C4, Rat — Daydy) + Cay ey Ra 7 — Daye) (2.18) ‘The auto- and cross-correlation function can be viewed as the superpositions of 280) and 057904) (one for each TDOA) type functions for ideal lowpass source signal and noises with flat bandlimited power spectral densities. The positions of the peaks in the auto- and cross-correlation function will not correspond exactly to the TDOAs due to the interference effect of the superpositions. The interference effect of the superpositions for two sinc functions is illustrated in Figure 2.12, where the dashed and dotted lines represent the two sinc functions #2 and S220 and the solid line indicates the sum of the two sine functions. The locations of the peaks in the sum are slightly shifted from those in the original sinc functions. The shifted positions of the peaks are denoted by Dé,, where subscripts ij indicates one of the six path pairs sidi,s2d2, didz, 8152, sida, dis and the superscript 6, which in this case is not an exponent, is included to indicate these positions introduce a bias on Dy; 33, 2 Figure 2.12: An Illustration of Shifted Peaks in the Sum (solid line) due to the Effect of Superpositions for Two sine Functions. The auto- and cross-correlation functions can be estimated with their time aver- ages. Estimates for the auto- or cross-correlation function are denoted fx(r), Rza(r) or Rj2(r), and given by Baal) = % ff zalt)zalt ~ rat (2.19) for m.n € {1,2} and sufficiently large T' to satisfy BT > 1, where B is the bandwidth of the signal s(t). If the TDOAs are sufficiently separated, they will produce distinct peaks in the auto- and cross-correlation functions and are said to be resolvable. Under these circumstances, the six TDOAs can be estimated from (2.19) [30]. ‘The models of the auto- and cross-correlation time difference of arrival (TDOA) estimators are shown in Figure 2.13 and 2.14 respectively. For the case of a signal and noise with flat low-pass spectra, these are minimum variance estimators [23]. ‘The ree ed signal is delayed by + and then multiplied by the original signal for 34 Pi) Peak Bi TDOA Fe (ae Locator Remover |B imates x(t) >} Delay r Figure 2.13: Model of the Auto-Correlation TDOA Estimator. pial TDOAs Peak Bias fs tia (ide Locator Remover Estimates Figure 2.14: Model of the Cross-Correlation TDOA Estimator. the auto-correlator or by another received signal for the cross-correlator. The time average of this resulting product gives the estimated auto- and cross-correlation fune- tions. A peak locator is used to locate the peaks of the multipath TDOAs in the estimated auto- and cross-correlation functions. In the case of the cross-correlator, the peak locator produces the locations of 4 peaks. The peak locator produces biased TDOA estimates, which pass through a bias remover to get the bias-corrected TDOA estimates. ~Stoa 00 goo ato 200 0200400 foo 800 1000 Range (neors) Figure 2.15: Six Hyperbolas Defined by Six True TDOAs for the Source at Depth 500 Meters and Range 500 Meters 2.5 Range and Depth Estimators In this study, the range and depth estimation for a given bearing is based on all of the six bias corrected TDOA estimates. Each of the six estimated TDOAs defines a hyperbolic line in the vertical y-z plane (Figure 2.8) that relates depth to range. Any pair of hyperbolas intersects at two points in the positive = half of the plane (i.e. beneath the ocean surface). Both points have the same depth and range but differ in dearing by 180°. Since only range and depth are of interest here, either of these points can be used to estimate the source location. If the received signals are noise free. the TDOA estimates are equal to the true TDOAs, and only two of the six hyperbolas are needed to determine the source location. Figure 2.15 plots the six noise free ayperbolas defined by the six true TDOAs for a source at a depth of 500 meters and a range of 500 meters. The solid (at bottom near the origin), dashed (at top near the 36 Depth (meters) "0 a a a 7) Range (meters) Figure 2.16: Six Hyperbolas Defined by Six Noisy TDOAs for the Source at Depth 500 Meters and Range 500 Meters. origin), dotted, dashdot, solid and dashed lines indicate the six hyperbolas. which are defined by six true TDOAs, Dads, Darts Dandss Desens Daydy and Dayag, respectively. Figure 2.15 shows that all six TDOAs produce distinct hyperbolas and all hyperbolas intersect at the true range and depth of the source. If the received signals are noisy, then, in general, the hyperbolas will not pass through the point of the true range and depth of the source. In this scenario. each pair of hyperbolas will, in general, intersect at a different point, but all intersections will be in the vicinity of the true range and depth. This is illustrated in Figure 2.16. which plots the six hyperbolas for the same two-receiver vertical array and the same source position used in Figure 2.15, but with additional noise corrupting the received signals. That is, the six hyperbolas are defined by six random variables, Ds,4,. Desdzs Dada» Ons: Daya and Deysqs and they are represented by solid (at bottom near the origin), dashed (at top near the origin), dotted, dashdot, solid and dashed curves, respectively. In this situation, the intersection of a single pair of hyperbolas will in general have more error than an estimate obtained by averaging, in some way, the range and depth given by the fifteen intersections, which are produced by the six pairwise noisy hyperbolas. ‘The equations for the six hyperbolas defined by the six, bias-corrected. TDOA estimates can be written in a matrix form as Sq=r, (2.20) where 4a -2 DR, ° 4b DE oO. = (1-H AD} -((s— 22)? - Abate) | (oon) (a-2P-8bR, (la 2)? ADR Mat (atn)?-2Dhg (lata)? - edhe) (ataP- eb, (ata edi, q= (xi and ct Dhan + 3 DE a (=f + (2.23) itd, ted tebe + bebe -lape (3 where c is the sound speed, and Dj; is the TDOA estimate of the delay path pair ag € { sudr, Seda, dida, $12, S1dzy disa}. ‘The vector q in (2.20) can be estimated using the least squares estimator 4 = (S7S)-'s’r. (2.24) Six Bias Corrected Multipath TDOA Estimates Dads Darts, Dandy Bans Dna and Darn xd Xs Least Squares Square Roots Solution = Receiver Depths a - x and 22 Figure 2.17: Model of the Range and Depth Estimators. =q+e, where E{e7e} This yields an estimate of q with error e = [e1 ¢2 3], ie. is minimum. ‘The range is estimated from the first element of 4 en Ze+ea]” by te = yx? te: ~ xe +Gh-. The depth can be estimated from either the second or the third element by 2, =\/z3 + ez = 2, + 7% for z, > e2 or #, = =, tes. It is pointed out that the solution obtained from (2.24) minimizes the sum of the mean squared errors in the estimates of q= (x? z? z,]". The error in the estimate of range is ££, which is considerably smaller than e1, as x, is much greater than e; for ranges of interest. The error in the depth estimate is either #& or e. If =, is much larger than es, then {2 should be smaller than es. With this reasoning the second element of q provides a better estimator of z, than does the third element. Therefore, for a submerged source, the second element of q is used to estimate the depth and the source position estimator is given by = [Va va] = [Vera vara)’. The model for the above range and depth estimators is shown in Figure 2.17. B= (ks 4] The range and depth estimators given above are computationally efficient as 39 the source range and depth can be estimated with a closed form equation (i.e. with- out iteration). It is claimed in [31] that the performance of this type of estimator approaches that of the maximum likelihood (ML) estimators. For this reason, the above range and depth estimators are chosen in this study. 2.6 Model Summary Models of the transmission system, two energy-based bearing estimators. and range and depth estimators are developed in this chapter. ‘The transmission system model encompasses: the source, the ocean acoustic channel and the localization system which consists of two vertically separated re- ceivers. The novelty of this model is the two receivers in the vertical array, each of which consists of cluster of directional sensors. The model for the ocean noise is a modified form of that introduced in [24]. The source model and the assumptions made for three main parameters in the ocean channel are similar to those made in (16 After establishing the model for the transmission system, two types of estimators are proposed: the two bearing estimators based on signal energy measurements and the range and depth estimators based on multipath TDOA information. The proposed two energy-based bearing estimators are novel. With this energy-based approach, the source bearing can be estimated from the average signal powers delivered by two directional sensors in one of the two clusters which can be supported by a single sonobuoy. The range and depth estimators are developed using the conventional TDOA approach. Models for the multipath TDOA estimators and the time delay 40 based range and depth estimators are presented to set the stage for the theoretical analysis to follow. 3. Performance of the Bearing Estima- tors This chapter evaluates the performance of the two energy-based bearing estima- tors proposed in Section 2.3. The performance of the bearing estimators is measured by their bias and variance. The bias can be removed if it can be theoretically pre- dicted. Therefore only the variance governs the accuracy of the bearing estimators The first section in this chapter develops expressions for the bias and variance of the two bearing estimators. Section 3.2 investigates the optimal values for the spread of the gain profile that minimize the bias and variance of the bearing estimators for 2 given angular separation. An example is given in Section 3.3, in which the com- parisons of the expressions for the bias and variance of each bearing estimator are made. In Section 3.4, the theoretical development is corroborated with Monte Carlo simulations. The chapter ends with a summary which forms Section 3.5. 3.1 Bias and Variance Expressions 3.1.1 Bias and Variance Expressions for Algorithm One The development of the expressions for the bias and variance of the bearing estimator is based on the model given in Figure 2.10. From (2.11), it can be seen that the bearing of the source is a non-linear function of the ratio of the two average signal powers. An approximation for the bias and variance of the bearing estimator 41 can be obtained in terms of the bias and variance of the ratio estimator by linearizing (2.13) using the first order Taylor series expansion. This approximation is valid when the bias of the power ratio estimator is much smaller than the true power ratio. ‘The derivation of the bias and variance of the bearing estimator is quite lengthy and involved due to the correlation between the noise voltages delivered by two sen- sors. This derivation therefore is given in Appendix A. For a large time-bandwidth product and a small bias on #1, the expressions for the bias and variance of the bear- ing estimator can be written in terms of the bias and variance of the ratio estimator and are given by (A.25) and (A.26) in Appendix A, which are E{6,-6} ~ Effi-n} (3.1) Anew = we) and Var {4.} : oe _ respectively. The bias of the ratio estimator, E {#1 — ri}, developed in Appendix A, (A21), is (ry ~ 1) 1 POR ag © leetert) + ages (F fe Etfii-n} —n1E {ee+1}), (3.3) and the variance of the ratio estimator, Var {7}, from (A.22) in Appendix A, is (ry — 1)? 2,2 Var {71} so (E {ected} — E {eneusi}) + slaps ((E fetes} EB {ex} E fextuss}) - 1 (E {eeeboa} ~ # {exes} E (ences) + 43 pares ((e et} - E {e,}) - 2ry(E {encues} ~ E (64) £ (cues}) +19 (E {chy} - B* {eus:})), (3-4) where pss is the mean value of source signal power (which includes the self-interference due to the surface-refiected path) that would be delivered if omnidirectional sensors were used for the receiver, c_ and €x41 are the noisy components of the powers deliv- ered by directional sensors k and k +1 due to noise, and E{e} denotes the expected values. More complete definitions of the symbols in Equations 3.3 and 3.4 can be found in Appendix A. In Appendix A, the noisy component of the average signal power delivered by each directional sensor, zg, is shown to consist of two parts: the cross product of the noise and source signal, denoted by ux, and the pure noise power contributed by noise itself, denoted by us. The first and second moments of uz and vg together with the mean value of 4s can be expressed in terms of the source signal power spectrum and the omnidirectional noise power spectrum for a large time-bandwidth product. The derivations of these moments are given in Appendix B. The four noisy components of the average signal powers delivered by the two sensors of interest are uk, ues, vg and visi. For a large time-bandwidth product, the two pairs ue, weer and vy, esi are approximately jointly Gaussian, i.e., ue, uke: ~ N(0,0; ug Fung Pnanegs) and ve, vig ~ (du Hogs Fs Fongs) Paaness > It is noted that Paynaz,> calculated using (2.8), is the correlation coefficient of the noises received at the two sensors. All moments of ¢x and x41 required in (3.3) and (3.4) can be expressed using the first and second moments of us, uss1, ve and veer and is so done in Appendix B. ‘The expression for the variance of the bearing estimator given in (3.4) is quite 44 lengthy and is a complicated function of the design parameters, such as the omnidi- rectional SNR, the spread in the gain profile and the angular separation of the look angles of two sensors. However, the third term in (3.4), which is a more straightfor- ward function of the design parameters, dominates for large omnidirectional SNRs and is in itself a good approximation. Thus for large SNRs the variance of the ratio estimator can be approximated by Var {71} wrap ((E a} - £ tet) - an (E {eutess} ~ E {ex} E {ee}) +7? (E {eb} £7 een}))- (3.5) ‘The expressions for the approximate bias and variance, given by (3.3) and (3.5) respectively, can be expressed explicitly in terms of the design parameters using the first and second moments of ui, west, vk and vigi. The resulting two expressions. (A.23) and (A.24), are given in Appendix A. An approximate bias and variance of the bearing estimator expressed in terms of the design parameters are obtained using (A.23) and (A ), respectively, with the results -¢ #0. - 0} = Eien = aN nT iPoamees 5 TC (2 Pramas x(i-( BT‘ 2x7(SNR)\ BT +) (3.6) and . rigs . Var {O} = a SRR CET (200 — 2M aman Hr) + ‘ 2 ge’ Brin (SNR) DraPeanses +1): (3.7) The restrictions placed on the design parameters in the development of (3.6) and (3.7) are summarized as follows: 1. The SNR is the ratio of total received source signal power (which includes the interference effects in a multipath environment) to total received noise power that would be experienced by an omnidirectional sensor located at the receiver. This SNR is assumed to be much greater than 1 2. Bis the bandwidth of the signal (the bandwidth may be wide or narrow and be either low-pass or band-pass in nature), and it is assumed that the time- bandwidth product, BT, is much much greater than 1. 3. Paanags iS the correlation coefficient of the noise voltages delivered by the two sensors. The value of the coefficient depends on the structure of the receiver and is determined using (2.8). In the development of (3.6) and (3.7). it was assumed that this coefficient is greater than 0.78 which restricts the spread in the gain profile to 7 < esi — He. Without significant degradation to the approximation, the bias expression given by (3.6) can be further simplified by ignoring the second term. This results in aa(eteae _ testy #a)(SNR) E{d,-9} ~ (3.3) Aig The variance of the beating estimator, given in (3.7), is further simplified without sacrificing significant accuracy by approximating the correlation coefficient, which is known to be greater than 0.78, with unity. This results in ites (ost _ teagty> Tae = mF GNRVBT) Var {6,} = 46 3.1.2 Bias and Variance Expressions for Algorithm Two The expressions for the bias and variance of the bearing estimator for Algorithm ‘Two is derived based on the model given in Figure 2.11. The procedure used is similar to that used for Algorithm One. The derivation of the bias and variance of this bearing estimator can be found in Appendix A with the results given by (A.33) and (A.34), which are . a : E{6,-6,} ~ eer tee ge (3.10) and : a 2 ver(i} = (Gopciyrcmn) Yer en where the bias of the ratio estimator, E {#2 ~ rz} from (A.29) in Appendix A, is given by : 1 | ethan) = ~ Grane” %} + 1 . ¢ 2) Ta ras (et 2 {B}), (3.12) and the variance of the ratio estimator, Var {#2} from (A.30) in Appendix A, is given by Veet = ray roeerag (© (8%) #8 wearer ((2(24} ) 1 a ty, wranserar pag ((E (4) 220) - 2re iG {& a} Biter) e{t}) +r iG {2} -E {)) (3.13) a7 Here 2, = £4 — x41 is the difference of the noisy components of the average signal powers delivered by two adjacent sensors, and 2, = ¢,+ék41 is the sum of their noisy components. For a large time-bandwidth product, the moments of 2 and & required in (3.12) and (3.13) can be expressed using the first and second moments of ia, test, vy and vest, which have the same meaning as in Algorithm One, i.e., they are the noisy components of the average signal powers delivered by the two sensors. The derivations for these moments are given in Appendix B. Since the moments of us, est, Uk and v4s1 can be expressed in terms of the power spectra of the source signal and the noise source, the resulting expressions for the bias and variance of the bearing estimator will be functions of the power spectra of the source signal and noise source. The lengthy variance expression, (3.13), can be considerably reduced. For a large omnidirectional SNR, the third term dominates so the other terms can be dis- carded with the result, Var {#2} weer (€ {ei} - £7 (e:}) - (9°(85) + 9°(O5**) 703 a }) +(e {2} -2*{2})). ean ‘As was done in Algorithm One, an approximate bias and variance for the ratio estimator, which are in terms of the design parameters, are found in Appendix A with final expressions (A.31) and (A.32). From these equations, the following expressions for the bias and variance of this bearing estimator are found: (1+ 9p) (es ae sty 4n?(uee1 — He)(SNR) 48 and Vee fi.) = 2" a —— (178) *(ueea — we)? (e+ eT) (SNR)(BT) seareee, myinsy +(a-90 Se) + eae areeeinnertea senate ee A +02, (a hanes) + 31+ Phames)) on arta (eS +e )swr) When BT > 1, (3.15) can be simplified to Ae (eta (ase) {6-6} = ~ “petites BNR) on without sacrificing much accuracy. This is the same expression as that for the ap- proximate bias of the bearing estimator for Algorithm One, (3.8) The variance expression given by (3.16) is further simplified without sacrificing significant accuracy by approximating the correlation coefficient with unity as was done for Algorithm One. This simplified variance is (3.18) which is the same as that for Algorithm One, given by (3.9). 3.2 Optimal Receiver Structure The simplified expressions for the bias and variance are identical for the two bearing estimation algorithms. Thus, the two algorithms yield approximately the same performance for the same set of parameters. As can be seen from (3.8) (which 49 is the same as (3.17)) and (3.9) (which is the same as (3.18)), the performance of both estimators depends on two receiver parameters: the spread in the gain profile, , and the angular separation between the look angles of two adjacent sensors, x41 — He This section finds the value of o that minimizes the absolute bias of the bearing estimator for a given angular separation, #1r41 — se, and the value of ¢ that minimizes its variance for a given jax¢i—x. These two values of c, which may differ, are referred to as optimal values with respect to bias and variance respectively. ‘The optimal values of ¢ are obtained by taking the derivative of (3.3) and (3.9) with respect to o, setting the resulting equations to zero and solving them. independently. ‘These optimal values will depend on the source arrival angle. The worst case performance in terms of both bias and variance is for an arrival angle 0, = ps1 Therefore it makes sense to seek the value of o that minimizes the variance or. if preferred, the bias, for the worst case arrival. The optimal value of o for this worst case, which is optimal with respect to bias, is given by nies © OSes — He) Os = Maen (319) The optimal value of o for the worst case with respect to variance is = 0.6(Hes — we); 8s = Hes (3.20) These are only approximate optimal values as they were obtained from approximate ions, (3.8) and (3.9). More accurate approximations for the optimal values of ¢ can be obtained using the more accurate expressions for the bias and variance, (3.6) and (3.2) for Algorithm 2 02] 0s] 04 os 2s} 03] Blas of Bearing Estimator (Radian) os 05] ot 2 as a ase Normalized Spread inthe Gain Profile Figure 3.1: Theoretical Biases of Two Bearing Estimators versus the Fraction <-*— when 0, = Has1- The Solid Line is for Algo- rithm One, while the Dashed Line is for Algorithm Two. One and (3.15) and (3.11) for Algorithm Two. As these expressions are quite lengthy. the solutions were found numerically. The plots of the bias and variance as a function of are shown in Figure 3.1 and 3.2. These plots were generated under the following conditions: an arrival angle 8, = 241, an angular separation of pesi — He = f radians, an omnidirectional SNR of 1.2 dB, flat low-pass source and noise power spectra with bandwidth B = 400 Hz, and an observation time of 1 second. ‘The solid lines in the two figures (Figure 3.1 and Figure 3.2) represent the bias and variance for Algorithm One, while the dashed lines in both figures indicate the bias and variance for Algorithm Two. Figure 3.1 shows that the optimal biases of the two estimators are very close, which correspond to 1, of Gepbies © eet ~ H&y for both algorithms. In Figure 3.2, both the minimal variances occur when Gaprar ~ 0.8 Variance of learng Estimator (Rains Squared) Dieses aris enact os eioesosrin4 [Normalized Spread inthe Gain Profile Figure 3.2: Theoretical Variances of Two Bearing Estimators versus the Fraction {4 when 0, = #i41- The Solid Line is for Algo- rithm One, while the Dashed Line is for Algorithm Two. (#k+1 — He). These optimal values differ from those given by (3.19) and (3.20) by 25 % and 33 %. However, the curves are very flat near the optimal points. both sets of optimal values result in virtually the same performance. This suggests the simplified expressions for the bias and variance are good approximations and produce valid optimal values. 3.3. Example Bias and Variance This section compares the expressions for the bias and variance of the two bearing estimators through a numerical example. The two adjacent sensors of interest are taken to have look angles zy = 0 radians and x41 = @ radians. This means that the angular separaticn between the look angles of the two sensors is equal to the spread in the Gaussian gain profile. This value of ¢ is very near the optimal value ‘eet Tadians. In this example, ¢ is taken to be for a worst case arrival angle of 8, § radians. That is, there is a total of M = 12 directional sensors in one cluster. The depth of the receiver is taken to be 200 meters below the ocean surface. ‘The source bearing is swept in steps of £ radians from —§ radians to % radians for four ranges: 490, 4,500, 8,500 and 12,500 meters. The depth of the source is held constant at 300 meters for all four sweeps. The power in the source is set so that the signal that would be received if an omnidirectional sensor was used in place of the proposed receiver, due to direct path only, is 20 dB above the noise power for a source range of 1,000 meters. The acoustic power radiating from the source is held constant, which means the power in the received signal changes with the range. For the source positions considered here, the received SNRs (for omnidirectional receivers), which include the effects of multipath, are 27.8, 9.8, 4.6 and 1.2 dB for ranges 490, 4.500, 8,500 and 12,500 meters. respectively. ‘The source and noise are assumed to be flat, ideal lowpass, broadband. sources with bandwidths B = 400 Hz. The observation time is 1 second. The bias and variance are plotted versus the normalized source bearing =, which means the look angles for the two sensors of interest correspond to abscissa values of 0 and 1. ‘The interval of interest is between zero and one. A log scale is used for the plot of variance. 03 = tance som aish ange 2200.0 . East 9 ca fags 12500 2 = = cos i i 5 Boo ost 7 a5 3 a5 7 is Normalized Source Bering Figure 3.3: Theoretical Bias of the Bearing Estimator for Algorithm One. The Lines Indicate the Results Calculated from (3.6), while the Mark “+” Represents the Bias Obtained using Simplified Equation 3.8 for All Four Ranges. 3.3.1 An Example for Algorithm One Figure 3.3 shows the bias results calculated using the full expression, (3.6). and the simplified expression, (3.8). Notice that the bias is zero at the abscissa value of 0.5, or 6 = $c radians, which is the point where the two sensors have equal gains. The bias calculated from the full expression is asymmetric about the equal gain point. Some asymmetry in the “full equation” curves is expected because an approximation was made in the theoretical development, which ignored a noise term in the denominator of the ratio estimator (see Appendix A). The bias obtained from the simplified expression is symmetric. This symmetry can be observed from (3.8). The comparison of the three variance expressions, (3.2), (3.7) and (3.9), is shown in Figure 3.4. This graph illustrates that the variance obtained using the full tos Vasiance of Bearing Estimator (Radian Squared) ‘Og ° os T 1s Normalized Source Beating Figure 3.4: Theoretical Variance of the Bearing Estimator for Algorithm One. The Lines Indicate the Results Calculated from (3.2), while the Marks *o” and “4” Represent the Approximate Variances Given by (3.7) and (3.9) Respectively for All Four Ranges, expression, (3.2), is asymmetric about the equal gain point. This is expected for the same reason that was given for the bias. Both the variances calculated using (3.7) and (3.9) are symmetric and give similar values in the interval of interest except for a range of 12,500 meters. The smallest variance also occurs at the abscissa value of 0.5. which is the equal gain point. The variance is approximately inversely proportional to the omnidirectional SNR. 3.3.2. An Example for Algorithm Two Figures 3.5 shows a comparison of the bias calculated using (3.15) and (3.17) ‘The comparison of the variance, calculated using (3.11), (3.16) and (3.18), is shown in Figure 3.6. The plots use the same convention as Figure 3.3 and 3.4. Both the Figure 3.5: 02 oas| on 00s} oy Bias of Bearing Estimator (Radian) 015} a2! 5 bs ° os T 15 Nonmulized Source Bearing Theoretical Bias of the Bearing Estimator for Algorithm Two. The Lines Indicate the Results Calculated from (3.15), while the Mark “+” Represents the Bias Obtained using Simplified Equation 3.17 for All Four Ranges. Figure 3.6: i 109] Range 490m Range 4500 m Range 8300 109 Range 12.500 a 1044 105] 104] 109} ‘Variance of Hearing Batista Radiuns Squred) og 7 os 7 15 Nomalzed Source Bearing Theoretical Variance of the Bearing Estimator for Algorithm Two. The Lines Indicate the Results Calculated from (3.11). while The Marks “o” and “+” Represent the Approximate Variances Given by (3.16) and (3.18) Respectively for All Four Ranges. bias and variance, calculated from all expressions, are symmetric about the equal gain point for this bearing estimator. For the interval of interest, (0, 1], the results obtained using the simplified expressions for this algorithm are closer to the “full equations” curves than those for Algorithm One. 3.4 Numerical Simulation The development of the expressions for the bias and variance of both bearing estimators involves a number of approximations. The principle approximation is the omission of a noise term in the denominators of each ratio estimator. To verify the validity of this and other approximations, Monte Carlo simulations were performed This section first describes the Monte Carlo simulation and then goes on to present comparisons of simulation results with theory. 3.4.1 Simulation Configuration A block diagram of the Monte Carlo simulation that was used for the corrobo- ration of the bias and variance of the two energy-based bearing estimators is shown in Figure 3.7. The Monte Carlo simulation consists of four parts as shown in Figure 3.7. The first part is the voltage signal generator, which is used to generate the voltage signals delivered by the two sensors of interest and is implemented as shown in Figure 3.8. The random number generator produces a white Gaussian sequence with zero mean and the appropriate variance. This sequence represents a sampled signal with sampling period T,, = 0.125 milliseconds, i.e. a sampling rate of fas = 8 kHz. This signal is filtered by an 8-order, Butterworth, lowpass, digital filter with Voltage Signal Generator teea(nTs] Ratio Estimators Fy or Fe Inverse Functions Bias and Variance Estimators i Estimated Bias Esti EX; - 8} & Figure 3.7: Monte Carlo Simulation Algorithm for the Bearing Estima- tors. ausgi Ranga Winbe ‘Generator Down Sample qT, Generic Sensor Signal (8) 9(95*") Additive Noise Additive Noise nalnT] reel Ts] ze[nTs] zeta [nT] Figure 3.8: Implementation of the Voltage Signal Generator for the Bear- 1g Estimators. 60 Integer Delay Fractional Delay ‘ Bl Delay Tsy ¥ Delay [oT ss Fo =1-Fp Figure 3.9: Implementation of the Linear Interpolator for Delayed Ver- sions of the Source Signal. its 3-dB-down point at 400 Hz to produce the source signal. The high sampling rate is used to facilitate the linear interpolation required to delay the source signal. The observation time is taken to be I second. ‘The time delays in the source signal are implemented with a linear interpolator after the source signal is filtered. The linear interpolator is implemented by two stages: one for an integer portion of the delay and another for a fractional portion of the delay as shown in Figure 3.9. The output of the linear interpolator is expressed mathematically by SnDoy —D] = (s(n = fp = Tu] = sll = Lo)Tal) Fo + sll ~ To) Tul (3.21) where [p is the integer part of 2 and Fp is the fractional part of #2 if D is the time delay of the signal. The delayed versions of the source signal are multiplied by the n coefficients calculated from (2.1) to form the direct and corresponding attenu: 61 surface-reflected path signals. A generic sensor signal is then formed by combining the direct path signal with the surface-reffected path signal and down sampling to a sampling rate of f, = 1 kHz (or the sampling period of T, = 1 milliseconds). The total signal voltages delivered by two sensors are formed by first weighting the generic sensor signal by the appropriate sensor gains and then combining them with the corresponding additive noises. The additive noise voltages delivered by the two sensors are produced by the additive noise generator, which is implemented as shown in Figure 3.10. V random number generators produce NV independent, identically distributed white Gaussian sequences with zero mean and the appropriate variance. The value of .V is set to 121 in the simulation. The sampling rate, f,, is 1 kHz. These noise sequences are filtered by NV, 8*-order, Butterworth, lowpass, digital filters with their 3-4B-down points at B = 400 Hz. The noise voltages delivered by the two sensors are formed by multiplying each of the NV filtered noise sequences by the appropriate sensor gain and then summing them. The second part of the simulation consists of the ratio estimators. The ratio estimators are implemented by first calculating the mean powers delivered by the two sensors. For Algorithm One, the ratio of the calculated powers is taken. For Algorithm Two, the ratio of the difference and sum of the calculated powers is taken. ‘The third part in Figure 3.7 consists of the inverse functions. This block converts the ratio estimates into the bearing estimates using (2.11) or (2.16) for Algori:hm One or Two, respectively. The output of the inverse block is a biased bearing estimate. which is denoted 6°. 9(68) n = Additive Noite - : ny[nT,) Gaussian e Generator 0 Sensor k Gaussian [Random Number Generator 1 Random Numbey-— "| Ha(w) A e : e e e = e 9(03**) a [Gaussian !Random Numbey Ha(w) Additive Noise riggs [nT] Generator V—1 Sensor kel (Oar) reece Figure 3.10: The Additive Noise Generator for the Bearing Estimators. 63 ‘The bias of the bearing estimators is estimated by subtracting the true value of the bearing from the estimated mean, i.e., he Ba) = EVM) - (3.22) where 64(i) is the estimate obtained from the i* trial in the simulation and N, is the number of trials used. This bias estimator, given by (3.22), is unbiased with variance Var{ 8,8: - 0.) = Yar), (3.23) where Var {0,} is the variance of the bearing estimators. ‘The estimated variance of the bearing estimators is calculated by 1 Ne San) 2 (3.24) 4 = ym (a0 - which is an unbiased variance estimator, i.e., Ef: If 68 is assumed to be approximately Gaussian, the variance of the above variance estimator is given approximately by (3.25) 3.4.2 Comparison of Simulation Results with the Theory ‘This section presents a comparison of the simulation based bias and variance with the theoretical calculations using the full bias and variance expressions for both bearing estimators. It is the same as the example given in Section 3.3 where the two sensors of interest had look angles jz = 0 radians and jug: = @ radians. As stated previously, this means that the angular separation between the look angles of the two sensors is equal to the spread in the Gaussian gain profile. Choosing o = 64 Hes1 — He tests the assumption made in the analysis that ¢ < pegs — He. Two cases are considered for each algorithm: § radians, and Beet ~ He 2. o = Mess — ue = & radians. As in Section 3.3, the depth of the receiver is set to be 200 meters below the ocean surface. ‘The source locations are also the same as those chosen in the example given in Section 3.3. The source bearing is swept in steps of ¢ radians from —$ radians to radians for four ranges: 490, 4,500, 8,500 and 12,500 meters. The depth of the source is held constant at 300 meters for all four sweeps. Again, for the source positions considered here, the received SNRs (for omnidirectional receivers), which include the effects of multipath, are 27.8, 9.8, 4.6 and 1.2 dB for ranges 490, 4,500, 8,500 and 12,500 meters, respectively. ‘The number of trials used for each source location is 50. These 50 trials are used to calculate the mean and variance of bearings for both algorithms. Both the bias and variance are plotted against the normalized source bearing. That is, the est hon scale of the abscissa in all of the plots is normalized to sensor separation, i.e., the look angles for the two sensors of interest correspond to abscissa values of 0 and 1. A log scale is used for the plots of variances. Algorithm One Figure 3.11 compares the simulation results with the theoretical calculation obtained from Equaticn (3.6) for the bias of the bearing estimator. The simulation 03; ++ Range 490 m ous Range 4500m 2 Range 8.500 m out + Range 12.500. hs ° OF T 1s Nonmaina Source Bearing Figure 3.11: Biases for Algorithm One for fist — Hk = 9 = Radians. The Simulation Results are Represented by the Marks. while the Theoretical Results from (3.6) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. results support the theory in the interval of interest, which is [0, 1, for all four ranges with perhaps the exception of a range of 12,500 meter, where is some disagreement for source bearings near zero. This indicates the approximations made to develop the theoretical bias are valid in the interval of interest at the omnidirectional SNRs encountered in the simulation. The variance of the bearing estimator is illustrated in Figures 3.12. The sim- ulation results support the theory in the region of interest for ranges 490 and 4.500 meters. However, differences start to emerge at a range of 8,500 meters. It is also noticed that the theoretical variance becomes increasingly pessimistic as the source bearing moves away from the interval of interest. This is due to 9*(9s*1) getting smaller which strains the assumption that 22(SNRus*@*) 5, 1. Figure 3.12: 66 tot 7 + Range 490 Te + Range 4300 : o Range 8500 m a _ 5 Range 12300 i 104 a e109 g S tos : # wo Hos o os T 1s Normalized Source Beating Variances for Algorithm One for esa — fe = o = § Radians. The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.2) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. 06 ++ Range 490m 004 + Range 4.500 © Range 8,500 . Range 12,500 m of Wearing Estimator (Radins) 8 8 004] 006 ‘Os ° os T 15 Nomalized Source Bearing Figure 3.13: Biases for Algorithm One for eyi ~ ux = 7 = % Radians. ‘The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.6) are Presented as Solid, Dashed, Dotted and Dashdot Curves for the Same Ranges. Figures 3.13 and 3.14. show the results for a system with twice as many sensors. ie. ket — Hk = @ radians, and with sensors that have a more directional gain profile. ie., 7 = jes — we = % radians. Again, with this choice of ¢ the assumption o < Hes — He is tested. Comparing Figure 3.13 with Figure 3.11 and comparing Figure 3.14 with Figure 3.12 shows that both the bias and variance of the bearing estimator are reduced significantly. For example, for a source at range 8,500 meters , beat with a bearing on the look angle of the second sensor in both cases, i. gs of #/6 radians and r/12 radians in 12 and 24 sensor receiver systems respectively, which correspond to abscissa value of 1 on the all graphs, the bias and variance differ by factors of about 4 and 8, respectively. 103 = wl’ + Range 490. Ee + Range 4500 8 é @ Range 8300 m 7 X Range 12500" i é po a 2 14 i a $0 in Wns 0 Os v 1s Normalized Source Bearing Figure 3.14: Variances for Algorithm One for ps1 —x = 0 = 3 Radians. The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.2) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. It can also be noticed that the theoretical curves are more accurate for o = =/12 radians. This is due to SNR) being a factor of two higher, which makes the assumption that it is much larger than I more appropriate. Algorithm Two. Figures 3.15 and 3.16 show the comparison of the simulation and theoretical results for the bias and variance of the bearing estimator for fry — Wk = 0 = = radians. The theoretical curves are noticeably different than the simulation results in the interval of interest for a range of 12,500 meters in the case of bias and 8,500 and 12,500 meters in the case of variance. Comparing Figure 3.15 to Figure 3.11 and comparing Figure 3.16 to Figure 3.15: ++ Range 490m Range 4.500. © Range 8.500. 2 Range 12,500 m © 20s} & & Bia of Hearing Psimaor (Rian) . -ous| hs ° 35 t 1s Normalized Source Bearing Biases for Algorithm Two for jeyi — we = The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.15) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. 69 ie al “+ Range 490m & o Rang 800 fo ieee Fos : i: : :7 g LLU 3 107] : aa, Normalized Source Bearing Figure 3.16: Variances for Algorithm Two for jure: —pe = o = Radians. The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.11) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. = oe | ar : ciescans z . x Range 12,500. m 17 ; i °F, 004} ‘Os o os 7 1s ‘Normalized Sour: Bearing Figure 3.17: Biases for Algorithm Two for wag1 — ue= 0 = % radians. ‘The Simulation Results are Represented by the Marks, while the Theoretical Results from (3.15) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. shows that both algorithms yield similar performance in the interval of interest. This is consistent with the simplified expressions for the bias and variance of the two algorithms, which are the same. Figures 3.17 and 3.18 show the theoretical and simulation results for x41 — (te = 7/12 radians = o radians. The differences between two chosen receiver systems are neatly the same as they were for Algorithm One for the same reasons. 103; 109| 104] os 104 109] woe Variance of Bearing Estimator (Radian Squared) “ox o os T is [Nomalize Source Bearing Figure 3.18: Variances for Algorithm Two for esi — se = 9 = 3 Radians. The Simulation Results are Represented by the Marks. while the Theoretical Results from (3.11) are Presented as Solid. Dashed, Dotted and Dashdot Curves for the Same Ranges. 3.5 Summary The horizontal angle of arrival of an underwater acoustic source can be esti- mated from a cluster of directional sensors supported by a single sonobuoy. The performance of the proposed algorithms does not depend on the spectral shape of the received signal. It depends on the ratio of total received source signal power to total received noise power and therefore can take full advantage of any high powered tonals in the signal. The analysis of such a system reveals that the bias and variance of error can be predicted theoretically. The variance of error for the two energy-based bearing estimators is approximately inversely proportional to the received omnidirec- tional SNR and the square of the number of the directional sensors, and proportional to the spread in the gain profile. The results of analysis show that for received om- nidirectional SNR of 9.8 dB, which could include the signal power in the tonals, the standard deviation of less than 0.0046 radians in bearing error can be achieved with 12 directional sensor elements. This standard deviation is reduced to 0.0006 radians when 2 directional sensor elements are used. 4. Theoretical Investigation of Correla- tion among the Multipath Time Dif- ference of Arrival Estimators The range and depth of an underwater acoustic source are estimated based on time difference of arrival (TDOA) information. The performance of the range and depth estimators is therefore closely related to the accuracy of TDOA estimators. which is studied in this chapter. The focus of this study in multipath TDOA es- timation is on the investigation of correlation among multipath TDOA estimators. First, the background of the multipath TDOA estimation is reviewed briefly in Sec- tion 4.1. Then, in Section 4.2, covariance expressions for multipath TDOA estimators are developed. An example for a two-receiver vertically separated array where the multipath TDOAs are estimated with two auto- and one cross-correlator is given in Section 4.3. The theoretical development given in Section 4.2 is corroborated with a Monte Carlo simulation. The details of which along with a comparison of the simu- lation and theoretically acquired results are given in Section 4.4. The chapter ends with a summary. 4.1 Background of TDOA Estimation As introduced in Chapter 1, TDOA estimation is a critical part of the conven- tional passive localization system and has been addressed extensively in the litera- 4

You might also like