I.

INTRODUCTION

Analysis of Queuing Behavior of Automatic ESM Systems

M. H. EL-AYADI Ain Shams University Egypt K. EL-BARBARY H. E. ABOU-BAKR Military Technical College Egypt

Radar electronic support measures (ESM) systems detect active emitters in a given area and determine their identities and bearings. The high arrival rate of radar pulses in dense emitter environments demands fast automatic processing of arriving pulses so that the ESM system can fulfill its functions properly in real time. Yet, the performance analysis of automatic ESM system in real life is difficult since both pulse arrivals and widths can be specified only probabilistically. The success of queuing theory in many applications such as computer communication networks and flow-control has encouraged designers to utilize queuing theory in qualifying and judging the performance of automatic ESM systems in dense emitter environments. The queuing behavior of these systems is analytically evaluated under different service disciplines and elaborate computer simulations validate the results. The analysis involves statistical modeling of arrival and departure processes as well as distribution of service times. It permits estimating the blocking probability due to high arrival rates of intercepted radar pulses or due to limited speed of the deinterleaver processor. Queuing analysis is shown to be quite useful to quantitatively assess tradeoffs in ESM systems design.

Manuscript received February 1, 2000; revised May 31 and November 2, 2000 and January 4, 2001; released for publication February 9, 2001. IEEE Log No. T-AES/37/3/08569. Refereeing of this contribution was handled by J. P. Y. Lee. Authors current addresses: M. H. El-Ayadi, 101 El Alamein Street, El Sohfaeyeen, Cairo, Egypt; K. El-Barbary, Electrical Engineering Department, Military Technical College, Cairo, Egypt; H. E. Abou-Bakr, Electrical Engineering Department, Ottawa University, Kingston, Canada. c 2001 IEEE 0018-9251/01/$10.00 1010

Early radar electronic support measures (ESM) systems relied on human operator interpretation of ESM receiver output to provide classification and identification of intercepted emitters. The steadily increasing density of radar pulse environments leads to the requirement of some form of automatic ESM processing to cope with high radar pulse arrival rates and to provide a real-time response. Automatic ESM systems consist of three main subsystems: 1) the receiver-encoder subsystem, that measures the parameters of each received radar pulse and encodes them by a digital word called pulse descriptor vector (PDV), 2) the deinterleaver (or the preprocessor) that rapidly sorts the PDVs into sequences each comprising a group of PDVs supposed to be emitted from the same radar, and 3) the main processor. The deinterleaver much reduces the data rate since it encodes every segregated sequence of PDVs by a single emitter descriptor vector (EDV). The main processor compares estimated EDVs with others stored in the threat library of ESM system in order to identify the type of intercepted radar. In some advanced ESM systems, the main processor can further from identified radar type, instantaneous position data supplied from onboard navigator, angle of arrival (AOA) information, and electronic order of battle (EOB) [1] stored in the threat library, determine the location of the detected emitter. The queuing behavior of the receiver-encoder and the deinterleaver in dense emitter environments is analyzed here. Queuing theory is concerned with the abstract mathematical modeling of systems subject to demands whose occurrences and lengths can, in general, be specified only probabilistically. Although, these systems are usually very complex, it is often possible to abstract from the system description a mathematical model whose analysis yields useful information about the quality of the service and the efficient utilization of the system. Here, reception of radar pulses, and extraction of their parameters as well as sorting of PDVs into separate chains is modeled as a finite-state machine operating as a queue with a single server. It is known that the application of the queuing theory is possible in principle if both the arrival processes at the queuing model input and the service time distribution inside it are statistically characterized. All queuing models discussed here are conventionally labeled as shown in Fig. 1 [4, 5].

Fig. 1. Queuing model labels. JULY 2001

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3

TABLE I Relation Between Arrival Rate of Radar Pulses at ESM Receiver Input and Sum of PRFs of Radars Illuminating ESM Receiver
N X i=1

= Number of Active Radars 7 8 9 10 11 12 13 14 15 16 17 18 19 measured [Pulse/s] 31485 33951 35441 38687 42439 47681 49388 52188 56362 58621 59857 60867 62860

PRFi

[Pulse/s] 31500 34000 35500 38750 42500 47750 49500 52250 56500 58750 60000 61000 63000

II. QUEUING MODEL REPRESENTATION OF ESM RECEIVER-ENCODER SUBSYSTEM A. Distribution of Interarrival Times In dense emitter environment, the interarrival time between successive pulses at the ESM receiver input is a random variable distributed according to a negative exponential distribution with parameter equal to the sum of the pulse repetition frequencies (PRFs) of all active radars in the instantaneous view of ESM receiver [6]. Since for negative exponential distribution 1= is the mean value (of interarrival time), then is also the average arrival rate of radar pulses. As the number and PRF diversity of active pulse emitters increase, the arrival process at the input of ESM receiver tends to be stationary Poisson process. That is given a time interval , the probability that exactly n pulses arrive at the ESM receiver input during is ( )n Pn ( ) = e : (1) n! The average arrival rate was measured through computer simulations of different number of active emitters with different PRFs.1 Both the calculated arrival rate and the measured one measured are provided in Table I. Clearly, the arrival rate of radar pulses at the ESM receiver input is directly proportional to the number of active radars illuminating the ESM receiver, and is equal to the sum of PRFs of these radars. The values of the interarrival time between successive radar pulses, intercepted in one observation interval of the ESM system were recorded. The moments of the interarrival time up to the fourth order were computed and denoted as Mim ; 1 im 4. The first four moments, Mit ; 1 it 4 of the negative exponential distribution with parameter
of simulation of instantaneous view of an ESM system are provided in Appendix A.
1 Details

where, kxk denotes the norm of the vector x. Small value of D 2 indicates closeness of measured distribution to the theoretical one. Thus, a smaller value of D 2 indicates a better fit. Even though small is good, the question how good a fit is given by a particular value of D 2 should be answered. Therefore, we assumed that the measured interarrival time could be expressed as a convex mixture of a random variate with the theoretical distribution and a normal random variate N (0, 1), accounting for the simulation and modeling imperfections. The mixture weights were for the random variate with the theoretical distribution and (1 ) for the normal variate, where 0 1. Thus, could be interpreted as a purity percentage. Calculating D 2 for the convex mixture as function of and in (29), we could establish correspondence between the values of D 2 and those of for given . This correspondence helped us interpret the significance of the values of D2 , obtained from computer simulations. The results are presented in Table II for different simulated arrival rates of radar pulses emitted from the active radars in the instantaneous view of ESM receiver. For the data in Table II, 0:26 106 D 2 5:90 106 which corresponds to 0:99979 0:99988. The third property, we verified, is that the number of received pulses at the ESM receiver input during any observation time is randomly distributed according to Poisson distribution with parameter . We simulated an instantaneous field of view with = 30000 pulse/s and observation time = 0:21 s.2 Then, we partitioned this observation interval into a number of subintervals N . We counted the number of pulses n occurring in each subinterval. We determined the frequencies of occurrence of different values of n. c ) where We calculated the least squares estimate (L L = 0:21=N and assuming a Poisson distribution model of measured frequencies. We applied the chi-square goodness-of-fit test [10]. We repeated the above procedure for different lengths L of the subintervals as shown in Table III. For the three cases considered in Table III, the least computed P -value is 0.1025. Hence, we would not reject, at the = 0:1 level of significance, the hypothesis that the number
2 The value 0.21 s was the largest observation time we could simulate for ESM system on the machine Pentium 166 MHz, given = 30000 pulse/s.

taken to be the sum of PRFs of simulated active radars were calculated. We defined a normalized squared distance D2 between the theoretical and the measured densities of the interarrival time as M1t M1m 2 M 2 , 1t M M M 2t 2t 2m (2) D2 = M3t M3m M3t M4t M4m M4t

EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS

1011

TABLE II Normalized Square Distance Between Theoretical Negative Exponential Distribution and Measured Distribution of Interarrival Times Between Received Radar Pulses at Different Arrival Rates measured [pulse/s] 31485 33951 35441 38687 42439 47681 49388 52188 56362 58621 59857 60867 62860 67590 M1m [105 ] 3.17 2.90 2.80 2.50 2.30 2.09 2.02 1.95 1.77 1.70 1.67 1.64 1.59 1.47 M1t [105 ] 3.17 2.90 2.80 2.50 2.30 2.09 2.02 1.95 1.76 1.75 1.66 1.63 1.58 1.47 M2m [109 ] 2.09 1.90 1.86 1.50 1.30 1.07 0.98 0.73 0.82 0.78 0.76 0.75 0.72 0.62 M2t [109 ] 2.05 1.70 1.50 1.33 1.01 0.87 0.82 0.73 0.63 0.57 0.55 0.53 0.51 0.43 M3m [1013 ] 2.10 1.90 1.89 1.50 1.16 0.83 0.81 0.79 0.66 0.63 0.63 0.61 0.60 0.51 M3t [1013 ] 1.90 1.70 1.34 1.03 0.78 0.55 0.45 0.42 0.33 0.29 0.27 0.26 0.23 0.19 M4m [1017 ] 2.80 2.60 2.50 1.30 1.40 0.99 0.95 0.90 0.81 0.78 0.76 0.75 0.72 0.65 M4t [1017 ] 2.40 1.70 1.50 1.00 0.73 0.46 0.39 0.32 0.23 0.20 0.18 0.17 0.15 0.11 D2 [106 ] 2.20 2.06 0.39 0.26 2.05 2.00 3.30 1.65 5.90 4.70 2.89 4.70 4.30 5.54

of pulses, received during given observation times, follows a Poisson distribution. B. Distribution of Service Times The distribution of service time inside the ESM receiver depends on what pulse parameters are measured and on the service discipline of the receiver-encoder subsystem [1, 2]. Generally, there are two service disciplines of an ESM receiver-encoder [1, 2], namely the paralyzable counter with constant dead time and the nonparalyzable counter. Both service disciplines do not allow waiting. Under the paralyzable service discipline, an ESM receiver-encoder [1, 2] can only process pulse which arrives after fixed time from the previous received pulse. The corresponding queuing model is the M=D=1=0 model3 [4, 5]. Typical values of fixed service time are the mean and the maximum widths of arriving pulses. Under the nonparalyzable counter service discipline an ESM receiver-encoder is ready to process a new coming pulse as soon as the previous pulse is expired. Thus the service time for each received pulse is equal to its width. For overlapping pulses the service time is the minimum of resultant width and certain maximum permissible value [8]. This avoids overloading by a CW signal. The distribution of the service time under paralyzable counter service discipline is given by (t ) and under the nonparalyzable counter discipline is given by the actual distribution of the widths of arriving radar pulses. A good statistical model of the service time is the Erlang distribution
3 The

with parameters and [4, 5], et f (t) = () t1 ! 1 E ft g = q 1 = E ft2 g E 2 ftg = p 1 =p : E ftg

(3) (4) (5) (6)

Notice that for = 1, the Erlang distribution reduces to the exponential distribution. This would be the distribution of the service time under the nonparalyzable service discipline when the main contributors to the arrival process at the input of ESM system are short pulse radars (like range-finders or tracking radars). For > 1, the mode of Erlang distribution shifts rightward from zero and the coefficient of variation =E ftg decreases. This might approximate the distribution of the service time under the nonparalyzable counter service discipline in dense environment of surveillance radars with low diversity of pulsewidths. Ideally, the distribution of widths of arriving pulses should be a multimodal distribution with the number of modes corresponding to the number of active pulse radar emitters. But assuming low diversity of widths and considering multipath and narrowband filtering effects the actual distribution might be a unimodal distribution. Finally, as ! 1, the Erlang distribution tends to the deterministic one (t 1=). Thus, the queuing model of an ESM receiver under nonparalyzable counter service discipline is M=G=1=0,4 [4, 5].
4 The

M=D=1=k model assumes that all service times have actually equal fixed values (deterministic values) and that we have a Poisson input process with fixed mean arrival rate, only one server and maximum number of customers allowed to wait is k . The symbol M indicates that the interarrival time is exponentially distributed, while the symbol D stands for a deterministic service pattern.

M=G=1=0 assumes that the queuing system has a single server, maximum number of customers allowed to wait is zero, and a Poisson input process (exponential interarrival times). The customers are assumed to have independent service times but with the same (general) probability distribution.

1012

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3

JULY 2001

TABLE III Chi-Square Goodness-of-Fit Test for Number of Received Pulses at ESM Receiver Input During Given Observation Times N = 2100, L = 100 s

The ESM receiver-encoder will be in the free state at time instant t + t if either of the events e1 or e2 happens. e1 : The ESM receiver-encoder is free at the instant t and no pulse arrives during the interval t. e2 : The ESM receiver-encoder is busy at time instant t, processing of current radar pulse terminates during the subsequent interval t and no new pulse arrives. From (1) we have Pn=0 (t) = e t P (e1 ) = Pf (t)e t Pf (t)(1 t): (8) (9)

j 1 2 3 4 5

b ) = 2:9669 Lease squares (L

Interval Nj f0, 1g f2g f3g f4g f5, 6, : : :g s= 430 479 500 347 344 = 4:68

Npj 428.71 475.65 470.40 348.91 376.33

Degrees of freedom = 5 1 1 P [2 > s] = 0:1969 N = 1050, L = 200 s

(Nj NPj )2 =Npj

j 1 2 3 4 5 6

b ) = 5:7992 Lease squares (L

Interval Nj f0, 1, 2, 3g f4g f5g f6g f7g f8, 9, 10, : : :g s= 159 160 184 165 158 224

Npj 178.54 149.93 173.90 168.08 139.24 240.31

In [9], it is shown that the probability P (e2 ) can be expressed in the form of 0 t, no matter the service discipline adopted and the distribution of the service time. Table IV provides expressions for 0 in four different cases. Therefore, we can express Pf (t + t) as (10) Pf (t + t) = Pf (t)(1 t) + 0 t which reduces to a first-order differential equation d P (t) + Pf (t) = 0 : dt f The steady state solution of the above equation is Pf (t) = 0 =: Substituting from Table IV into (10) we get the following cases. (12) (11)

Degrees of freedom = 6 1 1 P [2 > s] = 0:1311 N = 700, L = 300 s

(Nj NPj )2 =Npj = 7:09

j 1 2 3 4

b ) = 8:7811 Lease squares (L

Interval Nj f0, 1, : : : , 6g f7, 8g f9, 10g f11, 12, 13, : : :g s= 145 193 188 174 = 4:55

Npj 159.22 180.16 172.75 187.88

Case 1
P f (t ) = e

(13a) (13b)

Case 2 Pf (t) = Case 3

Degrees of freedom = 4 1 1 P [2 > s] = 0:1025

1 1 + w

(Nj NPj )2 =Npj

w Pf (t) = 1 + ewmin ewmax : (wmax wmin )

(13c)

III. EVALUATION OF FREE STATE PROBABILITY Pf (t) Under either service disciplines the ESM receiver-encoder is in one of the following two states. 1) Free State: where the ESM receiver-encoder is free and ready to receive any new coming pulse. The probability that the ESM receiver is free is denoted by Pf (t). 2) Busy State: where the system is busy processing a pulse and cannot receive and process another pulse. So, in this state the new coming pulse will be lost. The probability that the receiver is busy is denoted by Pb (t). Since the system must be in one of the above states, it is evident that P (7) f (t ) + P b (t ) = 1:

Case 4 Pf (t) =

(13d)

Fig. 2 compares the ratios of successfully processed pulses under paralyzable service discipline (case e) and under nonparalyzable service discipline (cases a, b, c, and d). Note that M=E1 =1=0 M=D=1=0 with = 1=. We assume that successfully processed pulses are only those completely received. Fig. 2 indicates that for a given arrival rate, as the parameter of the Erlang distribution increases, the ratio of successfully processed pulses under nonparalyzable service discipline decreases. But, for all finite it remains higher than the same ratio under paralyzable service discipline. In practice, there can exist dense emitter environments providing pulse arrival rates much higher than those presented in
1013

EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS

the pulses of the input flow which are blocked are not randomly selected. The departure process has therefore a residual effect and is not a Poisson process. The distribution of the interdeparture times of successive processed pulses fTd (l) is deduced from the interarrival times of the input flow of pulses as follows fthe proportion of the distribution of the interarrival times in the range (l, l + l)g fTd (l) l = fthe proportion of the distribution of the interarrival times in excess of g =
Fig. 2. Comparison of ratios of successfully processed pulses under paralyzable and nonparalyzable counter service disciplines, PW is distributed according to Erlang distribution with w = 1= = 12 micro seconds and for different : (a) = 1, (b) = 2, (c) = 5, (d) = 8, (e) = 1 (paralyzable).

el l , e

l> l

(14) (15)

fTd (l) = exp((l ));

Fig. 2. Clearly, at such very high arrival rates the ratio of successfully processed pulses will be extremely low. Possible cures are either to equip the ESM system with several receiving-encoding channels, (multiple-servers), or to limit the measurements of monopulse parameters to those that can be obtained from the pulse leading edge and imposing the paralyzable service discipline. In this case, the fixed service time is of order 0:1 s or even 0:05 s. From (13a), the ESM receiver-encoder will be able to process up to 106 pulse/s with success ratio 90%. The same conclusion can be derived from Fig. 2. The coordinates of the first point on curve e are: 1 = 31485 and success ratio r1 = 65:56%, given = 12 s. Clearly, exp(01 0 ) = [exp(1 )]c with c = 01 0 =1 . Thus, for 1 = 106 and 0 = 0:1 s, we 0 0:2647 = r1 = 89:43%. have r1 IV. ESM RECEIVER-ENCODER DEPARTURE PROCESS Under the paralyzable counter service discipline, the departure process is a renewal process since

where T d is the interdeparture time between output PDVs from the receiver-encoder subsystem and is the fixed service time of the receiver. We evaluated the squared distance between the theoretical and the measured densities of the interdeparture times, according to (2). The results are given in Table V for different simulated arrival rates, assuming = 11 s. For the data in Table V, 0:29 104 D 2 6:57 104 . Using (30), we interpret this result as 0:99911 0:99956. Now we are going to show that under nonparalyzable counter service discipline if the arrival process is a Poisson process, then so is the departure process. Yet it is necessary that the width of intercepted pulses be exponentially distributed or has an Erlang distribution with moderate . Suppose that n pulses arrive at the input of the ESM receiver during L seconds and only d of them are successfully processed by the ESM receiver-encoder subsystem and passed to the deinterleaver processor in the form of d PDVs. Thus, n d pulses arent processed or are missed. As indicated in the previous section each arriving pulse will with probability Pf find the ESM receiver in free state and with probability Pb in busy state. Hence, the number d of pulses successfully processed out of n pulses arriving during the interval L follows the

TABLE IV Expressions for 0 in Four Different Cases Service Discipline Case 1 Case 2 Case 3 Case 4 Note:
(1) w

Service Time Distribution Fixed value Exponential with mean value = w

(1)

Queuing Model M=D=1=0 M=M=1=0 M=E =1=0 M=G=1=0

0 e 1 + w w 1+ ewmin ewmax (wmax wmin )

Paralyzable Non-paralyzable Non-paralyzable Non-paralyzable

Erlang with parameter and mean w(1) Uniform in [wmin , wmax ](2)

(2) w min

is mean width of arriving pulses. and wmax are maximum and minimum widths of arriving pulses, respectively. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3 JULY 2001

1014

TABLE V Normalized Square Distance Between Theoretical Delayed Negative Exponential Distribution With Parameters and and Measured Distribution of Interdeparture Times Between Output PDVs at Different Arrival Rates measured [pulse/s] 31485 33951 35441 38687 42439 47681 49388 52188 56362 58621 59857 62860 M1m [105 ] 4.30 4.03 3.88 3.70 3.46 3.25 3.20 3.05 2.90 2.84 2.90 2.85 M1t [105 ] 4.27 4.04 3.90 3.68 3.45 3.19 3.12 3.01 2.85 2.80 2.77 2.65 M2m [109 ] 2.50 2.36 2.33 2.24 2.06 1.78 1.72 1.71 1.56 1.02 1.38 1.30 M2t [109 ] 2.25 2.15 2.13 2.02 1.74 1.46 1.38 1.27 1.19 1.07 1.04 0.97 M3m [1013 ] 2.73 2.02 2.10 1.74 1.52 1.27 1.29 0.92 0.85 0.80 1.13 1.05 M3t [1013 ] 2.70 2.21 1.90 1.50 1.24 0.93 0.85 0.79 0.62 0.61 0.53 0.47 M4m [1017 ] 3.49 2.68 2.58 1.75 1.27 1.19 1.08 1.04 1.01 0.97 0.88 0.75 M4t [1017 ] 3.44 2.60 2.40 1.60 1.17 0.78 0.69 0.57 0.44 0.38 0.36 0.31 D2 [104 ] 0.49 0.66 1.09 0.29 0.84 0.48 6.57 1.12 3.07 2.04 1.28 3.50

Note: Constant service time = 11 s. TABLE VI Normalized Square Distance Between Theoretical Negative Exponential Distribution With Parameter 0 and Measured Distribution of Interdeparture Times Between Output PDVs at Different Arrival Rates measured [pulse/s] 31485 33951 35441 38687 42439 47681 49388 52188 56362 58621 59857 60867 62860 67590 M1m [105 ] 4.07 4.01 4.02 3.60 3.30 3.05 2.99 2.88 2.69 2.65 2.63 2.62 2.60 2.45 M1t [105 ] 4.17 3.90 3.80 3.50 3.33 3.09 3.02 2.95 2.70 2.70 2.67 2.64 2.59 2.47 M2m [109 ] 2.50 2.55 2.50 2.10 1.79 1.45 1.39 1.30 1.18 1.14 1.20 1.16 1.14 1.02 M2t [109 ] 3.40 3.11 2.90 2.50 2.25 1.91 1.83 1.70 1.50 1.46 1.42 1.39 1.34 1.22 M3m [1013 ] 2.55 2.50 2.50 1.90 1.50 1.14 1.09 1.05 0.93 0.90 0.90 0.90 0.90 0.82 M3t [1013 ] 4.33 3.68 3.30 2.70 2.26 1.71 1.66 1.49 1.28 1.18 1.14 1.10 1.04 0.91 M4m [1017 ] 3.31 3.20 3.20 2.40 1.50 1.30 1.20 1.20 1.10 1.10 0.95 0.92 1.00 0.95 M4t [1017 ] 7.20 5.80 5.10 3.90 3.17 2.20 2.17 1.70 1.40 1.20 1.10 1.10 1.08 0.90 D2 [104 ] 5.90 3.16 2.40 0.74 0.05 5.80 0.93 11.00 8.00 9.00 0.31 0.02 0.82 0.65

Note: Nonparalyzable counter service discipline, average PW = 10 s.

binomial distribution, n Pfd Pbnd , Pd=n (L) = d

Substitution of (1) into (17) gives 0dn (16) where 0 = Pf : (19) From (18) it is clear that under the assumptions indicated above the departure process will be a Poisson process with departure rate 0 = Pf . Consequently, the distribution of the times between successive output PDVs from the receiver-encoder subsystem is a negative exponential distribution with parameter 0 = Pf . Computer simulations verified these conclusions and the results are presented in Table VI, where 0:017 104 D 2 11 104 , corresponding to 0:99920 0:99985. If the assumption made on the distribution of widths of arriving pulses is absent, i.e., if for example
1015

Pd (L) =

(0 L)d 0 L , e d!

d0

(18)

where Pd=n (L) is the probability that d PDVs will emerge from the ESM receiver-encoder subsystem given that n pulses are received at its input during time L. Let us recall that in dense environments the number of pulses arriving at the ESM receiver input is a random variable distributed according to Poisson distribution with parameter . The probability Pd (L) that d PDVs will emerge from the receiver-encoder in L seconds regardless of the number of arriving pulses, can be expressed using the law of total probability; i.e., 1 1 X X n Pd (L) = Pfk Pbnd Pn (L): Pd=n (L)Pn (L) = d n=d n=d (17)

EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS

0 < wmin PW wmax , then the values of d in (16)(18) should be restricted to the following ranges: in (16), [L=wmax ] d minf[L=wmin ], ng, and in (17), [L, wmax ] d [L=wmin ], where [x] denotes the integral part of x. Obviously in this case the departure process is no longer a Poisson process since (18) is not valid for all d, L 0. V. APPLYING QUEUING THEORY TO MULTIPLE-PARAMETER DEINTERLEAVING Computer-based deinterleaving algorithms are classified as interval-only algorithms and multiple-parameter deinterleaving algorithms. Interval-only algorithms operate on the interpulse-arrival times to estimate potential pulse repetition interval (PRI) values corresponding to the stream of PDVs coming out of the receiver-encoder subsystem. Every potential PRI value is used to segregate a repetitive sequence of PDVs, assumed to represent a single radar emitter. A well-known interval-only algorithm using sequential difference (SDIF) histograms is given in [7]. In dense emitter environment, (high pulse arrival rate and large number of interleaved radar pulse sequences), interval-only algorithms are of low efficiency and reliability. Multiple-parameter deinterleaving algorithms operate on the monopulse parameters. Usually, the AOA and the radio frequency are used to sort the PDVs into strings with similar primary parameters. Multiple-parameter deinterleaving is completed by PRI analysis of every segregated string, using SDIF histogram [7]. It is possible to use the pulse amplitude (PA) for later antenna-scan analysis, but this costs the deinterleaver a large amount of data storage. We focus here on the queuing behavior of multiple-parameter deinterleavers. In [1, p. 452], deinterleaving is defined as memory-intensive procedure converting PDVs into EDVs. A typical deinterleaving strategy for modern ESM system is described in [1, pp. 55, 56]. In Fig. 5 we present a possible implementation of this strategy. First, a frequency band coarse presort of arriving PDVs is done. Actually, each PDV is labeled with a frequency band number and is directed (through a demultiplexer) to one of N paths according to its frequency band label (N is the total number of frequency bands). In each path, a fine RF sort of arriving PDVs is carried out by attaching to every PDV a label indicating the serial number of the frequency bin comprising its RF. Suppose that every frequency band comprises M frequency bins. After the fine RF sort, a histogram of active RFs is constructed by counting the number of PDVs belonging to each one of the M frequency bins. PDVs, already served by the histogram of active RFs, are stored in a content addressable memory
1016

(CAM). The histogram of active RFs compares the count in every frequency bin with two thresholds. Once the count in any frequency bin exceeds the upper threshold, the presence of stable high-duty factor emitters (CW radars, pulsed-Doppler radars, and data-links) is declared. The complete label of this particular frequency bin is sent to the CAM to retrieve all PDVs with this label. They are sorted on the AOA, and the resulting strings are passed to PRI analysis. Simultaneously, the upper and lower limits of the frequency bin in which the count exceeded the high threshold are sent to an adaptive static filter (employing window addressable memory (WAM)) to block all new PDVs from the detected high duty factor emitters. This allows subsequent processing to be executed at a reduced data rate. After completing the histogram of active RFs, the label of every frequency bin with final count falling between the two thresholds is used to identify the PDVs, already stored in the CAM, and are belonging to frequency stable emitters. Again these PDVs are sorted on the AOA and the resulting strings are sent to PRI analysis. At last, the labels of the remaining frequency bins with final count below the lower threshold are used to retrieve from the CAM all the PDVs that are supposed to belong to frequency-agile emitters. They are grouped in one string, sorted on AOA, and then passed to PRI analysis. In all cases the nonagile components of the PDVs of every segregated sequence are averaged to encode the whole sequence by a single EDV. Let us now show how to apply queuing theory to the multiple parameter deinterleaving process discussed above. We start with the coarse presort. Let tcomp be the time needed to compare the frequency of a new arriving PDV with the upper and lower limits of one of the N frequency bands. If a sequential search is used, then at most N comparisons are needed to identify the band that matches the PDV being served. Hence, the service time in the coarse presort is a discrete random variable with probability density function (pdf) of the form bs (t ) =
N X i=1

p(i) (t i:tcomp )

(20)

where p(i) is the probability that P the incoming PDV matches the ith frequency band, N i=1 p(i) = 1. If the frequency bands are ordered and a binary search is used, then at most log2 N comparisons are needed. The service time is again a discrete random variable with pdf,
log2 N

bb (t ) =
0

X
i=1

p0 (i) (t i:tcomp )

(21)

where p (i) is the probability that the match is exactly Plog N found after i comparisons, i=12 p0 (i) = 1. If the
JULY 2001

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3

coarse presort is realized by employing a WAM, the frequency component of the new arriving PDV is simultaneously compared with the upper and lower limits of all frequency bands. Hence, the pdf of the service time will be bw (t) = (t tcomp ): (22)

tcomp < < T min where is the fixed service time of the receiver-encoder under the paralyzable counter service discipline and T min is the minimum of PRIs of detected high duty-factor emitters. It can be shown that the distribution of interarrival times of PDVs at the coarse presort can be approximated either to fI (l) = ( s)e(s)(l ) l (24)

If no high duty factor emitters have been detected, the adaptive static filter is not introduced in the path of the main stream of PDVs departing from the receiver-encoder. Then, with the latter under the paralyzable counter service discipline, the queuing model of coarse presort will be fTd =b=1=0, where fTd is given by (15) and the density b(t) is given by (20), (21), or (22). But with the receiver-encoder under the nonparalyzable counter service discipline and assuming exponentially distributed widths of intercepted pulses, the coarse presort queuing model will be M=b=1=0 with arrival rate 0 = =(1 + w), (from (13b) and (19)). In this case, expressions (23a)(23c) of blocking probabilities correspond respectively to the densities (20), (21), and (22) of the service time in the coarse presort. Sequential search pb = 1 Binary search
[log2 N ] N X i=1

p(i)e0 itcomp

(23a)

pb = 1 WAM

X
i=1

p0 (i)e0 i:tcomp

(23b)

pb = 1 e0 :tcomp :

(23c)

In order to estimate the primary speed requirement of the coarse presort, let us consider the following example: Pb = 10%, N = 8, 0 = 30000 PDV/s, p(i) = 1=N for 1 i N , p0 (i) = 1=[log2 N ] for 1 i [log2 N ] and 4 instructions=tcomp (2 subtractions to compare with RF upper and lower limits, 1 fetch, 1 load). With sequential search the processor will have to run at an instruction rate r 5:054 MIPS, with binary search r 2:258 MIPS, and with WAM r 1:139 MIPS. If N is increased to 16, the instruction rate should be raised above 9.528 MIPS for sequential search and above 2.817 MIPS for binary search. But for WAM, the instruction rate need not be increased. Once high duty-factor emitters are detected, the adaptive static filter is introduced before the coarse presort to censor all their subsequent PDVs from the main flow. It is necessary then to modify the distribution of interarrival times of PDVs at the coarse presort. Assume that both the number of detected high duty-factors emitters and their sum of PRFs, s, are small compared with the number and sum of PRFs of other emitters. Assume also that

with the receiver-encoder under the paralyzable counter service discipline, or to s [(s)=(1+w)](ltcomp ) fI (l) = l tcomp e 1 + w (25) with the receiver-encoder under the nonparalyzable counter service discipline and assuming exponentially distributed widths of intercepted pulses. Partitioning the flow of PDVs after the coarse presort into N subflows much alleviate the queuing problems in the subsequent processing phases. For example, the pdf of the service time in the fine RF sort is given by (20), (21), or (22) with N replaced by M , (the number of frequency bins), in (20) and (21) and according that a sequential search, a binary search or a WAM is implemented. Assume, just for informative purposes, that expressions (23a)(23c) apply to the fine RF sort.5 Assume also that the coarse presort blocks 10% of the arriving PDVs and those who come out are equally distributed among the N available paths. Thus, if the arrival rate at the coarse presort is = 30000 PDV/s, then for N = 8 we will expect an arrival rate at the fine RF sort, = 30000 0:9=8 = 3375 PDV/s. Now, for M = 16 and for the lowest instruction rate calculated above, (1.139 MIPS), we would expect Pb = 0:094 for sequential search, Pb = 0:029 for binary search, and Pb = 0:012 for WAM. Evidently, for higher instruction rates we have even smaller blocking probabilities. Furthermore, the service time in the histogram of active RFs is constant for all PDVs and is almost equal to 0:5tcomp which implies less blocking of PDVs than in preceding stages. For the AOA sort as well as for the PRI analysis the arrival pattern changes to batch arrivals at slow rate instead of single PDV arrivals at high rate. Thus, the AOA sort will almost cause no blocking of PDVs. The sizes of PDV batches (strings) arriving for PRI analysis are smaller than those of PDV batches arriving for AOA sort. Yet the service time for each string in the PRI analysis may vary significantly, because the emitters may be randomly agile in PRI and/or some PDVs in the string may be missing due to blocking in previous processing phases. From the above investigation of the queuing behavior, it has become clear that the frequency band
distribution of interarrival times of PDVs at the fine RF sort is no longer a negative exponential distribution.
5 The

EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS

1017

coarse presort is the true bottleneck of the studied deinterleaving process. The performance of the deinterleaver in terms of the blocking probability is improved when it is preceded by a prebuffer of size K . The function of this prebuffer is to store up to K of the arriving PDVs when the coarse presort is busy. However, if a new PDV arrives while the prebuffer is full, this PDV will be lost. The queuing model of the coarse presort in this case is M=b=1=K , assuming the receiver-encoder is under the nonparalyzable counter service discipline. Suppose that the deinterleaver is preceded by a prebuffer of size K . The instant at which the coarse presort completes servicing the nth PDV is the instant when the deinterleaver is ready to service the (n + 1)st PDV. Let us designate such instant by tn , n = 1, 2, 3, : : : , and the corresponding state of the prebuffer by the positive integer k , where k is the number of PDVs left in the prebuffer after the last processed PDV. Now we denote the probability that the prebuffer is at instant tn in state k by k (tn ), 0 k K . Assume that the input flow of PDVs is a Poisson process with average rate 0 . The probability that j PDVs arrive at the prebuffer input during the service time of PDV number n is Z 1 P (n = j , t) dt rj = P (n = j ) =
0

Fig. 3. Dependence of coarse presort blocking probability on the arrival rate of PDVs, constant prebuffer size = 5 PDVs, comparison time with one frequency band limits: (a) 45 s, (b) 35 s, (c) 25 s, (d) 15 s. N is total number of active emitters within instantaneous view.

P substitution. By imposing the constraint, K i=0 i = 1 we can determine the value of 0 and hence the values of the other state probabilities i , 1 i K . However, any analytic expression for the steady state blocking probability will be in expanded form that does not easily provide useful conclusions. VI. COMPUTER SIMULATIONS AND CONCLUSIONS

e0 t (0 t)j b(t) dt j!

(26)

where n is a random variable denoting the number of PDVs arriving at the prebuffer input while servicing the PDV number n and b(t) is given by (20), (21), or (22). The probabilities rj are transition probabilities associated with occurrence of different prebuffer states during servicing successive PDVs. It can easily be deduced that j (tn ) =
j X

i=[j=k ]

j i+1 (tn1 )rj ,

j = 0, 1, 2, : : : , K (27)

where [x] denotes the integral part of x. As n ! 1, the system composed of the prebuffer and the coarse presort reaches a steady state and the prebuffer state probabilities k , 0 k K become time independent. In particular, the steady state blocking probability will be given by the limit of k (tn ) as n ! 1. Moreover, the system (27) will take the form
j X

i=[j=k ]

j i+1 (i,1 ) = 0 j ,0 ,

j = 0, 1, 2, : : : , K (28)

where m,n equals 1 if m = n and equals zero for m 6= n. The above system can be efficiently solved for 1 , 2 , 3 , : : : , K in terms of 0 , using forward
1018

We simulated a receiver-encoder under the nonparalyzable counter service discipline, followed by a prebuffer of size K and a coarse presort into N frequency bands using sequential search. Figs. 3 and 4 show quantitatively the effect of the speed of the coarse presort processor and the size of the prebuffer on the blocking probability of the coarse presort at different arrival rates of PDVs. Fig. 3 shows that for a fixed prebuffer size and fixed average arrival rate of PDVs, the higher the speed of the presort processor, the lower the blocking probability. Also, the blocking probability is reduced by increasing the size of the prebuffer, for fixed processor speed as shown in Fig. 4. In particular, for the simulated emitter environment, when the single comparison time with RF limits of one frequency band is 35 s and the prebuffer size is 5 PDVs, then for arrival rates 20000 PDV/s the blocking probability is 50% (Fig. 3). Now, if the prebuffer size is increased from zero to 80 PDVs, the presort can process up to 23400 PDVs/s with blocking probability 50% (Fig. 4). Alternatively, if the single comparison time is decreased from 35 s to 25 s (by increasing the processor speed), the coarse presort with prebuffer size 5 PDVs can process up to 24000 PDVs/s with blocking probability 50% (Fig. 3). Thus in our case, 40% increase of processor speed is roughly equivalent to 16 times increase of the prebuffer size
JULY 2001

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3

Fig. 4. Dependence of coarse presort blocking probability on the arrival rate of PDVs, constant comparison time with one frequency band limits = 35 s, (a) prebuffer size = 0 PDVs, (b) prebuffer size = 20 PDVs, (c) prebuffer size = 40 PDVs, (d) prebuffer size = 80 PDVs. N is total number of active emitters within instantaneous view.

Fig. 5. Typical multiple-parameter deinterleaving scheme.

if the blocking probability is the measure of comparison. We could also assess another interesting tradeoff in design of coarse presort for simulated emitter environment. Fig. 4 indicates that at tcomp = 35 s, 0 = 27650 PDV/s, N = 8, and prebuffer size K = 80, the blocking probability is estimated to be 0.62. Meanwhile, expression (23c) indicates that the same value of blocking probability will be obtained when employing a WAM for coarse presort and removing the prebuffer. In this paper it has been shown that queuing models closely match experimental data and that

queuing analysis is quite useful to quantitatively assess tradeoffs in ESM systems design. For example, the number of bits used to generate the PDV will determine the needed comparison time inside the deinterleaver. Thus, the longer the length of the PDV, the higher the blocking probability of the deinterleaver. To decrease the comparison time, the number of bits used to represent the PDV should be decreased, but this degrades the resolution and accuracy of measuring the parameters of intercepted pulses. To maintain good measurement performance of each parameter without increasing the blocking probability of the deinterleaver we have either to increase the prebuffer size or to choose a faster processor to decrease the comparison time. Although, fast processors are expensive they ensure high average services rates, minimal reporting latency of intercepted threats and are usually associated with less hardware (no need for large buffers). The service time of the deinterleaver can also be decreased by optimizing the sorting algorithms and in some applications, by limiting the measurement of monopulse parameters in the ESM receiver-encoder only to those obtained from the pulse leading edge (short PDVs). Finally, the authors suggest the following extensions of this work: 1) development and evaluation of techniques for adaptation of service discipline to emitter environment; 2) further queuing analysis of design tradeoffs: single sever with limited waiting room (buffer) versus multiple servers, single parameter deinterleaver versus multiple parameter deinterleaver, high resolution of monopulse parameter measurements versus deinterleaver error performance; 3) queuing analysis of electronic counter measures (ECM) resource manager and
1019

EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS

in particular the effect of different priorities of demands. APPENDIX A A. Simulation of Instantaneous View of ESM System We assume that the radar emitters existent in the instantaneous view of an ESM system have technical parameters in the following ranges: PRF [16] KHz, PW [110] s, and RF [1.512.5] GHz. The AOAs of intercepted pulses vary from [5 71 ]. 2

B. Expressions of Squared Distances D 2 for Convex Mixtures 1) The theoretical distribution is f (l) = exp(l1 ), l 0 h
D 2 ( , ) = (1 )4 +

9 2 12 2 (1 )4 + (1 )2 + 3(1 )4 2 2 2 4 6 8 + 4 4 + 36 6 + 576 8 2

i2

(29) 2) The theoretical distribution is f () = exp((l )), l 2 2 : (30)

D2 (, , ) = 2

1 +

(1 ) + 9 (1 ) + 4

2 2 + 2 + 2

1 +

+ 6

+ 6 (1 )

6 6 3 2 + 3 + + 3 2

2 2 + 2 + 2 + 8

+ 3(1 )

24 24 12 2 4 3 + 4 + 3 + 2 + 4

Simulated radars are of low probability of intercept (LPI) and therefore have very low sidelobe level (SLL 40 dB). Thus the ESM receiver intercepts only mainlobe emissions from these radars. As we are simulating instantaneous views of the ESM system, the observation intervals are much shorter than the scan cycle of any of the simulated radars. Therefore, the individual radar scan rates are not considered in the simulations. The mainlobes of intercepted radars illuminate continuously the antenna of the ESM system during the relatively short observation interval (chosen to be 0.075 s). Thus, for an average pulse arrival rate of the order of 40000 pulse/s, the ESM receiver is expected to intercept about 3000 pulses/observation interval which is convenient for the storage capacity and the speed of the machine used for simulation (Pentium 166 MHz). Each intercepted pulse is represented by a PDV with components: RF, PW, AOA, and time of arrival of the pulse. We consider also both frequency-stable and frequency-agile radars. In every simulation frequency-agile radars are randomly selected from the set of active radars in the instantaneous view of ESM system. They are assigned agility bands falling completely within the frequency range of the instantaneous view of the ESM system (agility band = 510% of the RF). The pulses received from all radars during the observation time are sorted with respect to their time of arrival in ascending order. This results in a stream of randomly interleaved radar pulses simulating an instantaneous view of the ESM system.

REFERENCES [1] Curtis, S. (1986) Introduction to Electronic Warfare. Dedham, MA: Artech House, 1986. Davies, C. L., and Hollands, P. (1982) Automatic processing for ESM. IEE Proceedings, Pt. F, Comm., Radar & Signal Process, 129, 3 (1982), 146171. Tsui, J. B. (1986) Microwave Receivers with Electronic Warfare Applications. New York: Wiley, 1986. Hillier, S., and Lieberman, J. (1995) Introduction to Operation Research (6th ed.). New York: McGraw-Hill, 1995. Bunday, D. (1986) Basic Queuing Theory. London: Arnold, 1986. Bussgang, J. J., and Fine, T. L. (1963) Interpulse interval distribution in the environment of N periodic pulse radars. IEEE Transactions on Radio Frequency Interference, RFI-5 (1963), 710. Milojevic, D. J., and Provic, B. M. (1992) Improved algorithm for deinterleaving of radar pulses. IEE Proceedings, Pt. F, Comm., Radar & Signal Processing, 139, 1 (1992), 98104. Wilkinson, D. R., and Watson, A. W. (1985) Use of metric techniques in the ESM data processing. IEE Proceedings, Pt. F, Comm., Radar & Signal Processing, 132, 7 (1985), 221225. Abou-Bakr, H. (1998) Analysis of the queuing behavior of automatic ESM system in dense emitter environments. Master of Science thesis in Electrical Engineering, Military Training College, Cairo, Egypt, 1998. Low, A. M., and Kelton, W. D. (1991) Simulation, Modeling & Analysis (Industrial Engineering Series). New York: McGraw-Hill, 1991.

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

1020

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 37, NO. 3

JULY 2001

M. H. El-Ayadi received the B.S. degree (with distinction and honor) in electrical engineering from the Military Technical College, Egypt, the M.S. degree in electrical engineering from the VAAZ Academy, Czechoslovakia, the Ph.D. degree in signal processing and automatics from the University of Paris XI, France, in 1969, 1976, and 1980, respectively. He was a faculty member at the Military Technical College, Egypt in the Radar Engineering Dept. (19801981), then in the Electronic Warfare Engineering Dept. (19821996). There he became a Professor of Electrical Engineering (1991), Head of Electronic Warfare Engineering Dept. (1992), and Head of Electrical Engineering Depts. (1995). Since 1996, he has been a Professor of Signal & Systems at Ain-Shams University, Egypt, in the Computer Systems Dept. He supervised the inauguration of the Scientific Computing Dept. (19982000). His research interests include statistical signal and array processing and related applications in radar, communications, electronic warfare, distributed detection, and data fusion.

Khairy El-Barbary was born in Cairo, Egypt on November 22, 1958. He received the B.Sc. and M.Sc. degrees in electrical engineering from the Military Technical College, Cairo in 1981 and 1986, respectively. He received the Ph.D. degree in electrical engineering from the George Washington University, Washington, DC in 1994. Since 1994 he has been a staff member of the Electrical Engineering Department in the Military Technical College, Cairo, Egypt. His technical interests lie in the area of adaptive signal processing, interference cancellation, and array processing with applications to communications and radar systems.

Hussam Abou-Bakr was born in Cairo, Egypt on January 26, 1971. He received the B.Sc. and M.Sc. in electrical engineering from the Military Technical College, Cairo, Egypt in 1992 and 1998, respectively. During the period 1993 to 1999 he was a teaching assistant in the Department of Electrical Engineering, Military Technical College, Cairo. Currently he is preparing for the Ph.D. degree on electrical engineering at the Ottawa University, Kingston, Canada. He is interested in the area of signal processing with radar applications.
EL-AYADI ET AL.: ANALYSIS OF QUEUING BEHAVIOR OF AUTOMATIC ESM SYSTEMS 1021