0 оценок0% нашли этот документ полезным (0 голосов)

54 просмотров7 страницFeb 03, 2012

© Attribution Non-Commercial (BY-NC)

PDF, TXT или читайте онлайн в Scribd

Attribution Non-Commercial (BY-NC)

0 оценок0% нашли этот документ полезным (0 голосов)

54 просмотров7 страницAttribution Non-Commercial (BY-NC)

Вы находитесь на странице: 1из 7

Fig. 1. Radiometer block diagram. Several detectability models for the wideband radiometer have been published which allow easy calculation of the required input signal-to-noise ratio (SNR) for a given performance level. These models were developed because the exact solutions are intractable and must be evaluated using detection curves or numerical methods. The derivations of six common radiometer models are examined and comparisons for a variety of scenarios are provided. When possible, exact solutions are also compared. Results indicate that these models agree to within 0.5 dB for most cases of interest involving large time-bandwidth signals.

test statistic V exceeds the detection threshold V T (determined using one of a variety of criteria, such as Bayes, Minimax, Neyman-Pearson, etc. [9]), the signal of interest is assumed to be present. It is well known that if the input to the radiometer is strictly additive white Gaussian noise with two-sided power spectral density N0=2, the normalized random variable Y = 2V=N0 has a central chi-square distribution with = 2TW degrees of freedom [1, p. 298]: pn (y) = 1 2 =2 y (2)=2ey=2 , 2 y 0: (1)

I. INTRODUCTION In the design of low-probability-of-intercept (LPI) communication links, the performance of any potential intercept receiver must be considered. The performance of the interceptor is usually specified in terms of its probability of detection P , and d probability of false alarm, Pfa, and the required input signal-to-noise ratio (SNR). The most common intercept receiver is the wideband radiometer, which has been discussed extensively in the literature. To simplify the performance analysis of the wideband radiometer, several detectability models have been developed to easily determine the required input SNR for a desired performance level. The development of six such models is reviewed here and their results for a variety of performance requirements are compared. The models developed in Section II are those attributed to Torrieri [1], Edell [2], TEAL WING [3, 4], Engler [5], Park [6], and Dillard [7]. There are undoubtedly other models in use, but these are representative and easily found in the literature. Development of the exact solution, which is not solvable in closed form, is also included. In Section III, these models are then compared against each other, as well as exact results when feasible. II. MODEL DERIVATIONS A. Exact Solution The wideband radiometer is illustrated in Fig. 1. As shown, the detector measures the energy in a bandwidth W Hz over a time interval T s. If the

Manuscript received November 15, 1993; revised May 1, 1995. IEEE Log No. T-AES/32/1/00786.

If a signal with energy E (measured over T s) is present at the radiometer input, Y has a noncentral chi-square distribution with 2TW degrees of freedom and noncentrality parameter = 2E=N0 [1, p. 298]: psn (y) = p 1 y (2)=4 (y+)=2 e I(2)=2( y), 2 y0 (2)

where In (z) is the nth order modified Bessel function of the first kind. Fig. 2 shows an example of these probability density functions (pdfs) for the case where TW = 10 and = 20 (E=N0 = 10 dB). The performance of the radiometer, described in terms of its Pfa and Pd , is determined by integrating the conditional density functions as shown: Z 1 pn (y) dy (3) Pfa = Pd = Z

2V =N0 T 1

2V =N0 T

psn (y) dy

(4)

where V is the detection threshold against which V is T compared. A typical approach to performance analysis is to determine Pd and P for a given SNR and detection fa threshold. Other approaches would be to determine the required SNR for a desired Pd and Pfa, or determine the detection threshold based on an acceptable Pfa, and then solve for the Pd given an available SNR. Regardless of the method used, the integrals given in (3) and (4) are not solvable in closed form and must be evaluated numerically. Approximations and detection curves based on the numeric results have led to the development of receiver models which yield simple solutions. These models are discussed in the following sections.

467

B. Edells Model As the number of degrees of freedom becomes large (i.e., the time-bandwidth product is large), the chi-square and noncentral chi-square pdfs asymptotically become Gaussian by the central limit theorem. Edell used the Gaussian assumption to develop a simple, and commonly used detection model. In fact, several of the other models discussed in this work degenerate to Edells model under certain conditions. For TW > 100, Edell gives the following for Pfa and Pd [2]: Z 1 1 (x n )2 exp dx Pfa p 2 2n 2n n =Q (5) n Z 1 1 (x sn )2 Pd p exp dx 2 2sn 2sn sn =Q (6) sn

2 where n = 2TW, n = 4TW, sn = 2TW + 2E=N0, 2 and sn = 4TW + 8E=N0. The function Q(x) is the tail integral of the unit variance, zero mean Gaussian pdf: Z 1 x 1 2 1 ez =2 dz = 2 erfc p : (7) Q(x) = p 2 x 2

giving us a relationship between the desired performance and a required SNR, which is the normalized distance between the means of the signal-present and signal-absent pdfs. The value of d is obtained numerically or from curves given in [2, 10]. The term d2 is referred to as the detectability factor. Using the valuesp sn , n , and np for given earlier, we obtain d = E=N0 TW, or E=N0 = d TW. Finally, the required signal power to noise power spectral density (PSD) is obtained using S = E=T, yielding the following well-known detectability model: r S W =d : (11) N0 req T We refer to this model as the equal-variance Gaussian assumption (EVGA) model. For cases where TW < 100, the difference between the actual S=N0 using the exact chi-square statistics and that obtained using the EVGA model can be significant; for example, with Pd = 0:99, P = 1012, fa and TW = 1, the error is about 7 dB. To account for smaller TW products, Edell includes a correction factor which is defined as follows: = F(2 , Pd , Pfa, T, W) F(2 , Pd , Pfa, T, W) p = G(Gaussian, Pd , Pfa, T, W) d W=T

where F(2 , P , Pfa, T, W) is the predicted S=N0 using d accurate chi-square statistics for the specified Pd , Pfa, W, and T, and G(Gaussian, Pd , Pfa, T, W) is the value predicted using the Gaussian assumption. Curves for are given in [2, 10] for a variety of Pd , Pfa, and TW. With the correction factor, Edells complete model is r S W = d : (12) N0 req T C. Torrieris Model

The Q and erfc functions are 11 and invertible; hence, if y = Q(x) then x = Q1(y). Solving (5) and (6) for and equating yields = n Q1(Pfa) + n = sn Q1(Pd ) + sn Q1(Pfa) sn 1 n Q (Pd ) = sn : n n (8) (9)

Edell then used the assumption that sn n , implying that the SNR is small and the signal has little effect on the variance of the test statistic. This is generally true when dealing with LPI spread spectrum

468

Torrieri also uses Gaussian approximations for the chi-square distribution in the development of his detection model. His development follows that of Urkowitz [8], in which the signal and noise out of the bandpass filter of the radiometer are broken into their quadrature components. Sampling theorem notation is then used to facilitate the analysis, by approximating the integration of the radiometer with discrete summation of the in-phase and quadrature terms. For TW 1, these approximations become increasingly accurate. Torrieri shows that for large TW, EfVg N0TW + E varfVg N02 TW + 2N0E: (13) (14)

JANUARY 1996

Using the Gaussian assumption with no signal (E = 0), the false alarm probability is Z 1 1 (v N0TW)2 P q exp dv fa 2N02 TW 2N02 TW VT 1 V N0TW A T , Q@ q N02 TW 0

As TW ! 1, the second term in (20) approaches zero, resulting in the following: r r S 2W W 1 ( ) = [Q (Pfa) Q1(Pd )] N0 req T T (21) which is just Edells EVGA detectability model.

TW 1:

(15) D. TEAL WING Model Nicholson [3] presents detection models for several intercept receivers, which were originally developed by Bruce and Snow as part of the 1970s TEAL WING program [4]. As discussed previously, for TW > 100, the radiometer test statistic is essentially Gaussian, and the output SNR is related to the input SNR as follows [4]: (SNR)2 TW i (SNR)0 = (22) 1 + 2(SNR)i which for small (SNR)i simplifies to (SNR)0 = (SNR)2 TW i (16) (23)

Similarly, if the signal is present and perfectly aligned with the observation interval and bandwidth of the radiometer, then Z 1 1 Pd q 2(N02 TW + 2N0E) VT (v N0TW E)2 exp dv 2(N02 TW + 2N0E) V N0TW E A , Q @ qT N02 TW + 2N0E TW 1. 0 1

Note that the detection threshold used here is related to Edells by = 2V =N0. Solving (16) for the E=N0 T required to obtain a specified Pd with a given detection threshold yields E 2V 2 2 + T 2TW N0 N0 s 4V T 2TW, N0 (17)

which is typical of square-law receivers with small input SNR. The detectability factor d2 is used to relate the required input SNR to the desired performance, as in the case of Edells model: d2 = (SNR)0 = (SNR)2 TW i (24)

2 2 +

where d is given in (10). Solving for the input SNR yields d (SNR)i = p : (25) TW But (SNR)i is the ratio of signal power to noise power measured in the intercept bandwidth of the receiver, (SNR)i = S=(N0W), so the required S=N0 is identical to the EVGA model: r S W d : (26) N0 req T E. Englers Model Englers model [5] is based on Bartons detector loss function [11], which permits calculation of the required input SNR to achieve a given Pd and Pfa using the detection curves for a coherent receiver with TW = 1 (such as a single, unmodulated RF pulse). The detector loss function essentially converts a given amount of SNR, dI , which is available to a noncoherent receiver to an equivalent SNR, X0, which provides the same detection performance when applied to a coherent receiver. The general form of the detector loss function is [5] C(dI ) = b + dI adI (27)

p where = Q1(Pd )= 2. Solving (15) for the V =N0 T required to obtain the specified Pfa and substituting into (17) yields p E 2TW( ) + (, , TW), N0 p where = Q1(Pfa)= 2 and "r

p (, , TW) = 2 2TW

2

TW 1

TW 1 (18) #

8 2 2 +p 1+ 1 : TW 2TW

req

2W 1 ( ) + (, , TW), T T TW 1: (20)

CORRESPONDENCE

469

which gives the following mapping between dI and X0: to show that Parks model can also be expressed in Englers form: 2 dI adI q X0 = : (28) = 2 C(dI ) b + dI X0 + X0 + 18:4TWX0 S = (35) N0 req 4T Because X0 is the required SNR for coherent receiver performance, (28) is easily modified for use with any where X0 is defined in (32). time-bandwidth product: X0 =

2 TWadI : b + dI

(29)

G. Dillards Model Bartons detector loss function allows simple computation of the performance of a noncoherent receiver with any TW product using coherent detection curves for TW = 1. Urkowitz [12] used Bartons noncoherent integration formulas to show that receiver performance could be determined using noncoherent detection curves for TW = 1 as well. Using this approach, we can obtain the detection model given by Dillard [7]. Barton defines the noncoherent integration loss as Ln = Dn nDn = D1=n D1 (36)

Barton determined the coefficients a and b in the detector loss function by comparing the noncoherent and coherent detection curves in [9, pp. 297, 308], for a variety of detection requirements. For TW = 1, Barton found that a = 2:0 and b = 2:3 provided the best results. Engler showed that for large TW products (TW > 105), however, b = 2:0 was more accurate, and that the error incurred either way is less than 0.5 dB for all TW products. Solving (29) for dI with a = b = 2 yields q dI = X0 +

2 X0 + 16TWX0

4TW

(30)

Since dI is a ratio of signal power to noise power, we must multiply by the receiver bandwidth to get the required S=N0: q 2 X0 + X0 + 16TWX0 S = : (31) N0 req 4T Unlike the models discussed previously, this model is suitable for large and small TW products. The value of X0 depends on the desired performance (Pd and Pfa) and can be obtained from coherent detection curves in [9, p. 297]. It can also be evaluated numerically as follows: X0 = [Q1(Pfa) Q1(Pd )]2 = d2 : F. Parks Model Parks model [6], also suitable for all ranges of TW, is given as p S = d W=T (33) N0 req where the correction factor is s ! r 1 d2 TW = 1 + 1 + 18:4 2 : 4 TW d (32)

where Dn is the required single sample SNR for a desired Pd and Pfa when n samples are noncoherently integrated. Barton also shows that Ln = Cn (D + 2:3)=Dn = n C1 (D1 + 2:3)=D1 (37)

where Cn is the detector loss when the input SNR is Dn (note that the coefficient b = 2:3 from (27) is used). Equating (36) and (37) and solving for Dn yields q 2 4 2 D1 + D1 + 9:2nD1 (D1 + 2:3) Dn = : (38) 2n(D1 + 2:3) Dillards detection model is then obtained by completing the square in the numerator of (38) and using n = TW, = D1, and DTW = S=N0W, resulting in S 4:6W =p (39) N0 req 1 + 9:2TW( + 2:3)=2 1 which is Dillards equation (2) modified for S=N0 instead of E=N0. The value of = D1 can be obtained from single-pulse noncoherent radar detection curves or from an analytic approximation given in [9, p. 319]:1 q 1 2 ( 2 ln(P ) Q1(Pd ))2 : (40) fa Note that Dillards model can be modified using 2.0 instead of 2.3 in (37), resulting in S 4W =p (41) N0 req 1 + 8TW( + 2)=2 1

(34)

Hence, the Park and Edell models are identical in form, except curves for the correction factor are no longer required for small TW products. Park derives his expression for from Bartons detector loss function, using the coefficient b = 2:3 in (27) instead of b = 2:0 as used by Engler. Therefore, it is easy

470

SNR values obtained from curves and equations in [9] require adjustment by 3 dB, since DiFranco and Rubin use 2E=N0. JANUARY 1996

1 The

which, according to Engler, is more accurate for large TW. Like Torrieri and Edell, Dillard uses the Gaussian approximations for large TW as well. In fact, Dillards development proceeds identically to that of Edell. Using the normal approximations provided by Urkowitz [8], Dillard gives 2TW Pfa = F p (42) 4TW ! 2E=N0 + 2TW P =F p (43) d 4TW + 8E=N0 where is the detection threshold, and 1 F(x) = p 2 Z

x

which Dillard calls the full normal approximation. For 2E=N0 TW (small SNR), (45) can be simplified to E=N0 1 F (1 Pfa) P F p (46) d TW which Dillard refers to as the simple normal approximation. The simple normal approximation is equivalent to the EVGA model. This is easily verified by solving (46) for E=N0 and using F(x) = 1 Q(x) and F 1(y) = Q1(1 y). III. MODEL COMPARISONS Table I summarizes the radiometer detection models presented in this work. As seen in the previous section, they differ primarily in their assumptions and simplifications. Several models are based on the work of Urkowitz, who showed that the conditional pdfs for the test statistic of the radiometer are of the chi-square form, and that they become asymptotically Gaussian as TW becomes large. Other models are based on Bartons detector loss function and signal detection theory. In this section, we investigate how well these models predict the required SNR to achieve

471

ez

=2

Note that these equations are equivalent to (5) and (6). Solving (42) for and substituting into (43) gives 0 1 E=N p 0 F 1(1 Pfa) C B TW C (45) Pd F B p @ 1 + (2E=N )=TW A 0

CORRESPONDENCE

some desired performance level. Comparisons to exact results are also provided. In Fig. 3, results from the detection models are compared with exact solutions for a range of TW products, with Pfa = 106 and Pd = 0:9. FORTRAN algorithms from the International Mathematical Statistical Libraries (IMSL) [13] were used to evaluate (3) and (4) to obtain the exact results. The complete Edell model is not shown because of the inherent errors incurred in manually reading correction factors ( in (12)) from curves in [2, 10]. For TW = 100, the Dillard noncoherent integration and Park models are most accurate, while for TW > 100, the Torrieri and Engler models become more accurate. For TW > 105, the Torrieri, Engler, and EVGA 2 models converge to the exact results, while the Dillard and Park models give errors of 0.43 and 0.3 dB, respectively. The S=N0 errors will also vary with Pd and Pfa. This is illustrated in Fig. 4, in which Pfa = 103 and TW = 1000 are used. As shown, Dillards model provides a nearly constant error for a broad range of

2 The

Pd , while Englers model provides the smallest error for all Pd . Torrieris model is accurate for large Pd , but gives a comparatively large negative error for Pd < 0:1. At Pd = 0:5, the Torrieri and EVGA models are equivalent. Fig. 5 shows receiver operating characteristic (ROC) curves for Pfa = 106 and TW = 10. Again, we see that Dillards noncoherent integration model is the most accurate for small TW, although all of the models agree within about 0.3 dB of the exact results. The EVGA and Torrieri models were not used since they are not appropriate for small TW. Fig. 6 shows ROC curves for P = 108 and TW = 109. In this case, fa the Torrieri, EVGA, and Engler models provide the same results, represented by the solid line. The Park and Dillard models differ from the EVGA curve by about 0.3 and 0.4 dB, respectively. Exact results were not calculated for the very large TW case because of limitations in the IMSL software, but from Fig. 3, it is reasonable to assume that the EVGA model produces nearly exact results. At this point we might wonder what we have really accomplished, since these detectability models are not expressible in closed form. For example, all of

JANUARY 1996

472

the models require computation of the inverse Q function. Computation of Q however is significantly less involved than recursively solving (3) and (4). Furthermore, most mathematics software packages include built-in variants of the Q function, such as erfc(x) or erf(x). They often include mechanisms for computing inverses of these functions, as well. Hence, as we see from the results in Figs. 36, the radiometer models presented here provide reasonable results, and are significantly easier to use than exact computations. IV. CONCLUSIONS

[8]

[9]

[10]

[11]

[12]

Several detection models for the wideband radiometer have been presented. The purpose of these models is to provide a simple means of predicting the [13] required SNR to achieve some desired performance (Pd and P ). Comparisons with exact results showed fa that the Torrieri, Engler, EVGA models converge to the exact results for very large time-bandwidth products. For TW > 1000 the maximum error using any of the models is less than 0.5 dB. Converting Earth-Centered, Earth-Fixed Coordinates to Geodetic Coordinates

ROBERT F. MILLS Department of Electrical and Computer Engineering Air Force Institute of Technology Bldg. 640, Suite 218 2950 P Street Wright-Patterson AFB, OH 45433-7765

Urkowitz, H. (1967) Energy detection of unknown deterministic signals. Proceedings of the IEEE, 55 (Apr. 1967). DiFranco, J. V., and Rubin, W. L. (1980) Radar Detection. Dedham, MA: Artech House, 1980. Woodring, D., and Edell, J. D. (1977) Detectability calculation techniques. Report NRL 5480, Naval Research Laboratory, Washington, DC, Sept. 1977. Barton, D. K. (1969) Simple procedures for radar detection calculations. IEEE Transactions on Aerospace and Electronic Systems, AES-5 (Sept. 1969). Urkowitz, H. (1973) Closed-form expressions for noncoherent radar integration gain and collapsing loss. IEEE Transactions on Aerospace and Electronic Systems, AES-9 (Sept. 1973). International Mathematical Statistical Libraries (1991) FORTRAN algorithms. International Mathematical Statistical Libraries, Houston TX, 1991.

An improved algorithm to convert Earth-centered, Earth-fixed (ECEF) coordinates to geodetic coordinates is presented. It is

faster than other methods and has accuracy equal to that of one GLENN E. PRESCOTT Department of Electrical Engineering and Computer Science of the most accurate existing algorithms, for heights greater than University of Kansas 3,000,000 m. Lawrence, KS 66045 REFERENCES [1] Torrieri, D. J. (1992) Principles of Secure Communications Systems. Boston: Artech House, 1992. Edell, J. D. (1976) Wideband, noncoherent, frequency-hopped waveforms and their hybrids in low-probability of intercept communications. Report NRL 8025, Naval Research Laboratory, Washington, DC, Nov. 8, 1976. Nicholson, D. (1988) Spread Spectrum Signal Design: LPE and AJ Systems. Rockville, MD: Computer Science Press, 1988. Bruce, J. D., and Snow, K. D. (1974) Final technical report TEAL WING program. Probe Systems report PSI-ER327, Probe Systems, Inc., Sunnyvale, CA, Oct. 1974. Engler, H. F., and Howard, D. H. (1985) A compendium of analytic models for coherent and non-coherent receivers. Technical report AFWAL-TR-85-1118, Air Force Wright Aeronautical Laboratory, Sept. 1985. Park, K. Y. (1978) Performance evaluation of energy detectors. IEEE Transactions on Aerospace and Electronic Systems, AES-14 (Mar. 1978), 237241. Dillard, R. A. (1979) Detectability of spread-spectrum signals. IEEE Transactions on Aerospace and Electronic Systems, AES-15 (July 1979), 526537.

I. INTRODUCTION An earlier algorithm [1], which computes an initial estimate then refines it, is improved in three ways: 1) An error at high latitudes is eliminated by using acos ( ) instead of asin ( ), 2) terms are rearranged to simplify the computation of sin( 0 ), and 3) a more accurate height formula is used. The new algorithm is derived in the next section. To estimate accuracy, a testbed program is used to determine the maximum and average output errors of the algorithm over a large number of input test points generated by the testbed program. Each geodetic test point is exactly converted to Earth-centered, Earth-fixed (ECEF) coordinates, approximately converted by the algorithm to geodetic, and exactly converted back to ECEF where magnitude of the 3-dimensional error is computed. The testbed program

Manuscript received November 4, 1994; revised May 10, 1995. IEEE Log No. T-AES/32/1/00787. U.S. Government work, U.S. copyright does not apply.

[2]

[3]

[4]

[5]

[6]

[7]

0018-9251/96/$5.00

CORRESPONDENCE

473

## Гораздо больше, чем просто документы.

Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.

Отменить можно в любой момент.