Parametric Clutter Rejection for Space-Time Adaptive Processing A. Lee Swindlehurst and Peter Parker Brigham Young University Dept. of Electrical & Computer Engineering Provo, UT 84602 voice: (801) 378-4343 fax: (801) 378-6586 email: swindle@ee.byu.edu A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Motivation: Motivation: Circular Circular Array Array STAP STAP • Small sample-support interference rejection techniques needed for non-stationary clutter • Offset need for larger sample-support due to increased clutter rank • Use small sample-support plus parametric modeling to reduce computation A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Data Model • M antennas, N pulses • Target in primary range bin p x p (t ) = ba(θ )e jωt + c p (t ), spatial steering vector • t = 0, 1, 2, L, N − 1 clutter, jammer, noise, etc. Space-Time Slice [ X p = x p (0 ) x p (1) L x p (N − 1)] = ba(θ )v T (ω ) + C p [ temporal Mω/7 V = 1 H steering vector A. Lee Swindlehurst Department of Electrical & Computer Engineering L H M (1 −1) ω/7V Brigham Young University ] Provo, UT 84602 Data Model (cont.) • Vectorized Forms 1. [ S = YHF ( ; S ) = E Y ( ω) ⊗ D ( θ ) + F S 2. [ S = YHF ( ; 7S ) = E D (θ ) ⊗ Y (ω ) + F S • Secondary Data target-free range bins A. Lee Swindlehurst {F N } N = 1, L , 1 V N ≠S ( (F N ) = 0 , ( (F N F N* ) = 5 Department of Electrical & Computer Engineering interference covariance Brigham Young University Provo, UT 84602 Space-Time Autoregressive Modeling L z ) = ∑ H i z −i −1 • Define i =0 M ‘ x M matrices • Model: for some L, ( z −1 )ck (t ) = H0ck (t ) + H1ck (t − 1) + L + HLck (t − L) = k (t ) is spatially and temporally white • To estimate (z −1 ) , solve Ns min H 0 ,L , H L A. Lee Swindlehurst closed-form least-squares solution N ∑ ∑ k =1 i = L +1 ( z )ck (i ) −1 Department of Electrical & Computer Engineering 2 Brigham Young University Provo, UT 84602 Least-Squares Solution • Define H * = [H 0 L H L ] • Then Ns N k =1 i = L +1 ∑ ∑ c k (N ) ck (L + 1) Ck = M L M ck (1) ck (N − L ) ( z −1 )c k ( i ) 2 [ = H * C1 L C Ns ] 2 • Solution: columns of H are M’ least dominant singular vectors of: [C A. Lee Swindlehurst 1 L C Ns ] M(L+1) x Ns (N-L) matrix Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Prior Work • Vector AR models used previously for clutter modeling by Michels, Rangaswamy, etc. • Used to develop GLRT and AMF detectors for Gaussian and non-Gaussian assumptions • Here, we use AR model to generate structured subspace orthogonal to clutter subspace • Clutter is then projected out of primary data prior to space-time filtering A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 An Interpretation STAR filter attempts to minimize clutter power: dimension M‘(N-L) x 1 k 0L 0 HL HL−1 L H0 O M c =+ c =0 O k k 0 M O O L L 0 0 H H H 0 L L−1 Least squares algorithm finds block Toeplitz +that minimizes Ns ∑ k =1 2 k Ns = ∑ + ck 2 k =1 Span(+ ) orthogonal to clutter subspace if it dominates white noise: R = + ⊥Q + ⊥ + clutter & jamming A. Lee Swindlehurst Department of Electrical & Computer Engineering 2 I white (thermal) noise Brigham Young University Provo, UT 84602 Filtering the Primary Data Clutter lies in span of + , so we project it out: ( ) x p = + ++ + xp 123 / banded block Toeplitz Like an eigencanceller, but here subspace is structured. Occasionally, a more computationally efficient approach is: • Use SVD to find orthogonal complement of +: ++ ⊥ =0 dimension MN x (M-M’)N+M’L • Filter data as follows: ( x p/ = I − + ⊥+ A. Lee Swindlehurst Department of Electrical & Computer Engineering ⊥ )x p Brigham Young University Provo, UT 84602 Algorithm Summary [ 1. Use SVD on secondary data to solve for H0 H1 L HL computational order: 2. Form +and filter data: ( ) ] O Ns M 2 (L + 1) (N − L ) 2 P+ x p ( ) computational order: O M / M 2 (L + 1)2 (N − L ) 3. Perform regular beam and Doppler filtering for detection computational order: negligible −1 Resultant test statistic is (v ⊗ a ) P+ * x p not (v ⊗ a ) R x p * A. Lee Swindlehurst Department of Electrical & Computer Engineering * Brigham Young University Provo, UT 84602 STAR STAR Filter Filter Example Example 18 element ULA, 18 pulses SJR = -40dB SCR = -40dB SNR = 10dB Fully adaptive STAP requires ~60 secondary snapshots for similar performance A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Jammer Jammer Obscured Obscured Target Target 18 element ULA, 18 pulses SJR = -40dB SCR = -40dB SNR = 10dB Fully adaptive STAP requires ~70 secondary snapshots for similar performance A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Computational Comparison Some typical numbers: M = 20, N = 18, L = 2, M / = 15 • Proposed approach: O(57,600Ns ) + O(864,000) • Fully adaptive algorithm: ( ) ( ) O Ns M 2N 2 + O M 2N 2 = O(130,000Ns ) + O(9,700,000) rank of clutter covariance ≅ 75 • PRI staggered algorithm: ( ) ( O N s M 2K 2 N + O # of sub-CPIs = 3 A. Lee Swindlehurst ) M 2K 2N = O (65,000Ns ) + O( 2,900,000 ) rank of sub-CPI covariance ≅ 45 Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Model Order Selection When correctly specified, structure of +implies clutter rank equals MN − M / (N − L ) = 360 − (18 − L ) M / Approach for simulations: 1. Pick “arbitrary” value for M / penalty term 2. Choose L using MDL: L = arg min p Ns ∑ k =1 (p ) k 2 M / M (p + 1) − M / (M / + 1) / 2 + log2 Ns MN prediction error when L = p For M / = 15, L = 1-2 for long ranges, 2-3 for short ranges A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Model Order Sensitivity 50 km Range 300 km Range 20 16 18 M’=5 M’=10 M’=15 M’=20 M’=30 12 Average SINR loss (dB) Average SINR loss (dB) 16 M’=5 M’=10 M’=15 M’=20 M’=30 14 14 12 10 8 6 10 8 6 4 4 2 2 0 0 2 4 6 8 10 12 14 0 0 L A. Lee Swindlehurst Department of Electrical & Computer Engineering 2 4 6 8 10 12 L Brigham Young University Provo, UT 84602 14 Sample SINR Loss 50 km Range, Training length 30 samples 30 20 SINR (dB) 10 0 −10 Optimum STAR PRI FULL −20 −30 −40 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 Normalized Doppler A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Average SINR Loss 25 20 15 SINR (dB) 10 5 Ave. SINR loss is area between curves 0 −5 −10 −15 −20 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 Normalized Doppler A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Performance at Short Ranges 20 km Range 50 km Range 20 20 M’=15 M’=20 PRI FULL M’=15 M’=20 PRI FULL 16 Average SINR loss (dB) Average SINR loss (dB) 18 15 10 14 12 10 8 6 4 2 5 1 10 2 10 0 1 10 2 10 Training length Training length No jammer present A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Range-Varying STAR Filter • To improve performance at short ranges, use linearly varying matrix coefficients: Ns H i (k ) = H i + k − + i , k = 1,L, Ns 2 coefficient i for secondary snapshot k • Analogous to ESMI technique of Hayward • Solution found from SVD of A. Lee Swindlehurst C1 L Ns 2 L ( )C 1− 1 CNs −1 ( )C Ns −1 2 Department of Electrical & Computer Engineering Ns −1 CN s Ns CNs 2 Brigham Young University Provo, UT 84602 Performance with Range-Varying Weights 30 km Range L=4 30 km Range L=3 12 12 ESMI−−15 ESMI−−20 STAR−−15 STAR−−20 PRI FULL 10 Average SINR loss (dB) Average SINR loss (dB) 10 8 6 4 ESMI−−15 ESMI−−20 STAR−−15 STAR−−20 PRI FULL 8 6 4 2 2 0 1 10 2 10 0 1 10 2 10 Training Length Training Length No jammer present A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Performance at Long Range 300 km Range 8 7 M’=15 M’=20 PRI FULL Average SINR loss (dB) 6 5 4 3 2 1 0 1 10 2 10 Training length A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Performance vs. Range with Jammer 20km Range 300km Range 40 20 18 M’=15 M’=20 PRI FULL 16 30 25 20 15 M’=15 M’=20 PRI FULL 10 5 Average SINR loss (dB) Average SINR loss (dB) 35 14 12 10 8 6 4 2 0 1 10 2 10 0 1 10 Training data length 2 10 Training data length JCR = 10 dB, Jammer DOA 30 deg., SINR Loss at 0 deg. A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Effect of Sample Support Training length 30 samples 20 18 18 M’=15 M’=20 PRI FULL 14 12 10 8 6 14 12 10 8 6 4 4 2 2 0 50 100 150 200 250 M’=15 M’=20 PRI FULL 16 Average SINR loss (dB) 16 Average SINR loss (dB) Training length 100 samples 20 300 350 0 0 50 Range (km) 100 150 200 250 300 350 Range (km) No jammer present A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 400 Performance vs. Sample Support with Jammer 100 secondary data vectors 30 secondary data vectors 25 M’=15 M’=20 PRI FULL 20 Average SINR loss (dB) Average SINR loss (dB) 25 15 10 5 0 M’=15 M’=20 PRI FULL 20 15 10 5 0 50 100 150 200 250 300 350 0 0 Range 50 100 150 200 250 300 350 Range JCR = 10 dB, Jammer DOA 30 deg., SINR Loss at 0 deg. A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 3-D STAR Filter for Hot Clutter Idea: add matrix taps to exploit correlation in fast-time: P min ∑ Ns N ∑ ∑ p =0 k =1 i =L +1 p ( z )ck − p (i ) −1 2 For example, with P=2, solution involves SVD of C2 L CNs −1 CNs C1 L CNs −2 CNs −1 Low computation, sample support make full 3-D STAR solution feasible for larger P as well A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602 Hot Clutter Examples 30 km Range L=2 30 km Range L=1 30 30 3D−−15 3D−−20 STAR−−15 STAR−−20 3D−−15 3D−−20 STAR−−15 STAR−−20 25 Average SINR loss (dB) Average SINR loss (dB) 25 20 15 10 20 15 10 5 5 0 1 10 2 10 0 1 10 Training Length A. Lee Swindlehurst Department of Electrical & Computer Engineering 2 10 Training Length Brigham Young University Provo, UT 84602 Conclusions • STAR filtering ideal for STAP problems that require small secondary sample support • Structured model yields computationally efficient solution • Simulations with realistic circular array data show promising performance • Easily extended to handle hot or range-varying clutter models A. Lee Swindlehurst Department of Electrical & Computer Engineering Brigham Young University Provo, UT 84602