• purdy, blankenship, muehe, stern, rader, and williamson

8/14/2019 Purdy, Blankenship, Muehe, Stern, Rader, And Williamson

1/24

PURDY, BLANKENSHIP, MUEHE, STERN, RADER, AND WILLIAMSONRadar Signal Processing

V O LU ME 12 , N U MB E R 2 , 2 0 00 L I NC O LN LA B OR ATO RY JO UR N AL297

Radar Signal ProcessingRobert J. Purdy, Peter E. Blankenship, Charles Edward Muehe,

Charles M. Rader, Ernest Stern, and Richard C. Williamsons This article recounts the development of radar signal processing at LincolnLaboratory. The Laboratorys significant efforts in this field were initially drivenby the need to provide detected and processed signals for air and ballistic missiledefense systems. The first processing work was on the Semi-Automatic GroundEnvironment (SAGE) air-defense system, which led to algorithms andtechniques for detection of aircraft in the presence of clutter. This work

was quickly followed by processing efforts in ballistic missile defense, first insurface-acoustic-wave technology, in concurrence with the initiation of radarmeasurements at the Kwajalein Missile Range, and then by exploitation of the

newly evolving technology of digital signal processing, which led to importantcontributions for ballistic missile defense and Federal Aviation Administrationapplications. More recently, the Laboratory has pursued the computationally challenging application of adaptive processing for the suppression of jammingand clutter signals. This article discusses several important programs in theseareas.

of techniques developed for one mission area to othermission areas. For example, Lincoln Laboratorys ef-forts on air defense were applied to the needs of airtraffic control, satellite communication contributedto developments in space surveillance, and speechprocessing and solid state physics both contributedsignificantly to radar signal filtering. Particularly sig-nificant have been the pathfinding efforts in digitalsignal processing, and the successful application of this field to many important problems across variousareas of application.

The SAGE Air-Defense SystemIn the early 1950s Lincoln Laboratory participated inthe first application of digital technology to radarsignal processing. The Semi-Automatic Ground En-vironment (SAGE) Air Defense System was under de-velopment, and there was a need to transmit targetinformation from the radars over narrow-bandwidthtelephone lines to the direction centers. The solutionto this problem was the sliding-window detector il-lustrated in Figure 1. The name sliding window refersto the short length of time that a rotating antennas

T signal processing at Lin-coln Laboratory had its genesis in research ef-forts undertaken at the MIT Radiation Labo-

ratory during World War II [1]. These efforts, along with similar efforts at Bell Telephone Laboratories[2, 3], provided a theoretical foundation for many important developments in signal processing at many organizations during the ensuing years [4]. With theformation of Lincoln Laboratory in 1951, this theo-retical foundation was initially applied to programs inair defense. Soon, however, the stringent needs of bal-listic missile defense required the application of bothsignal processing theory and practice. Subsequently,signal processing requirements from fields as diverseas air traffic control, space surveillance, and tacticalbattlefield surveillance also stimulated the develop-ment and implementation of powerful new signalprocessing techniques and technology.

The essence of signal processing is its combinationof theory, efficient computational algorithms, and theimplementation of these algorithms in hardware.One interesting aspect of the history of radar signalprocessing at Lincoln Laboratory is the transference


2/24


298 L IN C OL N LA B OR ATO RY J OU RN A L V O LU ME 1 2, N UM B ER 2, 2 00 0

FIGURE 1. The sliding-window detector, operating withideal signal input. (a) The binary-quantized video signal afterthe application of an initial threshold to the range gate of in-terest. (b) The accumulation of the binary count of succes-sive returns from the range gate during one radar-beam-width traversal time. (c) The resulting binary sequenceshowing detection of the target when the count exceeds an-other threshold . The beam-split estimate of azimuthal po-sition corresponds to the midpoint of the interval duringwhich the cumulative sum exceeds the threshold [5].

beam dwells upon a target. Each implemented rangegate was assigned an accumulator. In each range gatethe video output from each radar pulse was sampledand subjected to an initial threshold. This output was

assigned a 1 value and added to the accumulator if the initial threshold was exceeded. A 0 value meantno detection and 1 was subtracted from the accu-mulator. The accumulator was never allowed to gobelow zero. A target was declared when the sum in theaccumulator exceeded a second threshold, as shownin Figure 1(b), and the end of the run was declared

when the sum in the accumulator fell below a thirdthreshold. The midpoint between these declarations

was generally used as the azimuth estimate of the tar-get, as shown in Figure 1(c). In the absence of a target

the receiver noise would normally cause the accumu-lator sum to hover well below the second threshold.The sliding-window detector approximated what ahuman operator would do in deciding on the pres-ence of a target on a radar plan-position-indicator

(PPI) display, and produced approximately the samenoncoherent integration gain as does the human op-erator. For each detection a single digital word con-taining range, azimuth, and strength of target was as-

sembled and sent over the telephone line. Analyses of the performance of the sliding-window detector werereported by Gerald P. Dinneen and Irving S. Reed[6]. The sliding-window detector, which was later re-named the common digitizer, became the standardmethod for detection in long-range ground-basedsurveillance radars for both air traffic control andmilitary applications.

Ballistic Missile Defense With the increasing ballistic missile threat in the

1950s, the Laboratory became heavily involved in de-veloping signal processing technology to address theincreasingly sophisticated radar signals that were usedto make measurements on ballistic missile reentry complexes. The theoretical basis for radar signal de-sign was advanced by the application of radar ambi-guity-function analysis, especially in high-clutter en-vironments [7, 8]. The problem then became one of identifying the appropriate technology for hardwareimplementation. Initial efforts used commercially available technology and were of limited capability [9]. Fortunately, technology was advancing, and twoapplication areas that were unique to the Laboratory proved to be particularly successful: surface-acoustic-

wave signal processing and digital signal processing.

Surface-Acoustic-Wave Signal Processing In the late 1960s a number of researchers around the

world became interested in the potential use of sur-face acoustic waves (SAW) for providing new types of compact filters that could operate in a frequency range from a few tens of hertz up to a few gigahertz.

Among other applications, the projected device pa-rameters seemed well matched to implementing ana-log pulse-compression filters for radars. As a result,the development of SAW devices for military use be-gan in several laboratories. One of the earliest efforts

was established at Lincoln Laboratory under the lead-ership of Ernest Stern [1012]. In the late 1960s thisgroup began to pursue the development of SAW de-vices for radar and communications applications.

16

12

8

4

(c)

(b)

(a)

2016 24 28 32 40 44 4836

Pulse number

0

1

0

8 12

0

1

Azimuth, time


3/24



The First Reflective Array Compressors The challenge for SAW technology was to achievesufficiently precise devices with the right combina-

tions of correlation time and bandwidth to be usefulin radar systems. The earliest SAW dispersive delay lines for use as radar-pulse compressors employed ametallic pattern of interdigitated electrodes depositedon the surface of a piezoelectric crystal such aslithium niobate [13, 14]. The electrodes launched anacoustic wave on the crystal surface; the electrode pat-tern was arranged so that it would be responsive tothe specific received signal. This interaction yieldedthe desired chirp response. The approach worked rea-sonably well for bandwidths of a few tens of mega-

hertz and for time-bandwidth products of one hun-dred or less, but it failed to yield sufficiently preciseresponse and low sidelobes when tried at higher time-bandwidth products.

Results obtained previously with some low-bandwidth acoustic filters [15] suggested to the LincolnLaboratory SAW group that reflection of SAWs from

arrays of grooves etched into the crystal surface couldyield a more nearly ideal device response than thatobtained with metallic electrode arrays. To explorethis hypothesis, the SAW group realized that experi-

ments were needed to elucidate the physics of surface wave reflections, new technology was needed to litho-graphically define and etch the reflective arrays, andnew device models and design techniques had to bedeveloped.

A great deal of the technological groundwork forthis process was established during 1971. By 1972,fabrication of the first reflective-array compressor(RAC) was initiated; this device is illustrated in Fig-ure 2. The first RAC device was a linear-FM filter

with a 50-MHz bandwidth (on a 200-MHz carrier)

matched to a 30- sec-long waveform [1618]. Thisarrangement yielded a time-bandwidth product of 1500, more than an order of magnitude greater thanthat achieved by interdigital-electrode SAW devices[19]. The response was remarkably precise; the phasedeviation from an ideal linear-FM response was only about 3 root mean square (rms). Pairs of matchedRACs were used in pulse-compression tests in whichthe first device functioned as a pulse expander and thesecond as a pulse compressor. The compressedpulsewidths and sidelobe levels were near ideal.

Armed with these encouraging results, researcherstook the next step by developing RAC devices for spe-cific Lincoln Laboratory radars.

RAC Pulse Compressors for the ALCOR Radar The ARPA-Lincoln C-band Observables Radar, or

ALCOR [20], on Roi-Namur, Kwajalein Atoll, Mar-shall Islands, had a wideband (512 MHz) 10- sec-long linear-FM transmitted-pulse waveform (see thearticle entitled Wideband Radar for Ballistic MissileDefense and Range-Doppler Imaging of Satellites,by William W. Camp et al., in this issue). ALCOR

was a key tool in developing discrimination tech-niques for ballistic missile defense. The wide band-

width yielded a range resolution that could resolve in-dividual scatterers on reentering warhead-like objects.This waveform was normally processed with theSTRETCH technique, which is a clever time-band-

width exchange process developed by the AirborneInstrument Laboratory [21, 22]. The return signal is

Output transducer (SAW to signal)

Metal filmof varying width

Etched grating

Inputtransducer(signal to SAW)

Etched grating

FIGURE 2. A phase-compensated reflective-array compres-sor, or RAC. The input transducer converts an electrical sig-nal into a surface acoustic wave (SAW) that propagates

along the surface of the crystal. The grating etched into thecrystal reflects the wave at a position determined by the in-put frequency and the local spacing of the grooves in thegrating. High frequencies reflect close to the input trans-ducer, while low frequencies reflect at the far end of the grat-ing. A second reflection sends the SAW to the output trans-ducer, where it is converted back into an electrical signal.The desired delay versus frequency is set by the geometry ofthe device. Deviations from the desired response can betrimmed out by a metal film of varying width deposited onthe device.


4/24



FIGURE 3. The ALCOR all-range wideband analog pulsecompressor developed jointly by Lincoln Laboratory andHazeltine Laboratory.

FIGURE 4. RAC sidelobe performance in compressing a10- sec 512-MHz-bandwidth pulse. (a) The compressedpulse and its sidelobes on a 1-GHz carrier frequency, shownon a linear scale. (b) The envelope of the compressed pulseand its sidelobes on a logarithmic scale of approximately 6dB per division. The horizontal scale on both graphs repre-sents 5 nsec per division.

mixed with a linear-FM chirp and the low-frequency sideband is Fourier transformed to yield range infor-mation. For a variety of reasons, the output band-

width and consequently the range window were lim-

ited. For example, the ALCOR STRETCH processoryielded only a thirty-meter data window. Therefore,examination of a number of reentry objects, or thelong ionized trails or wakes behind some objects, re-quired a sequence of transmissions.

This sequential approach was inadequate in deal-ing with the challenging discrimination tasks posedby reentry complexes, which consist not only of thereentry vehicle, but also a large number of other ob-

jects, including tank debris and decoys, spread outover an extended range interval. What was needed

was a signal processor capable of performing pulsecompression over a large range interval on each pulse.Lincoln Laboratory contracted with Hazeltine Labo-ratory to develop a 512-MHz-bandwidth all-rangeanalog pulse compressor employing thirty-two paral-lel narrowband dispersive bridged-T networks built

out of lumped components, to cover the bandwidth.The resulting processing unit, shown in Figure 3, waslarge (it filled about seven relay racks) and complex,and it required a great deal of tweaking to yield rea-sonable sidelobes. Cost and complexity loomed large

when plans were made for a series of reentry tests in which matched pairs of pulse compressors would berequired. In a parallel effort, the Lincoln Laboratory SAW device group was challenged to develop pulsecompressors that could meet the all-range needs of

ALCOR. This task would mean extending the band-

(b)

(a)


5/24



width of SAW RAC technology by an order of magni-tude, which would increase the time-bandwidthproduct well beyond that achievable with any existinganalog device technology.

During 1972 and 1973, Lincoln Laboratory devel-oped a 512-MHz-bandwidth (on a 1-GHz interme-diate frequency [IF]) 10- sec RAC linear-FM pulsecompressor [23]. In ALCOR, an active circuit withfeedback generated the linear-FM chirp, so that theRAC devices were to function as all-range pulse com-pressors matched to that waveform. To suppress rangesidelobes, a Hamming window was built into theRAC devices by varying the etch depth of the groovesas a function of position.

Midway in the development effort, significant dif-

ficulty was encountered in achieving sufficiently pre-cise amplitude and phase responses. Subtle litho-graphic and etching effects yielded errors in groovedepths and positions that measured only a few tenthsof a nanometer, but these very small errors were largeenough to degrade the compressed-pulse sidelobessignificantly. A trimming technique was developed toachieve an adequately precise response. This tech-nique required measuring the device and the subse-quent deposition of a corrective metal pattern of vary-ing width on the crystal surface of the RAC, asillustrated in Figure 2. The resulting precision al-lowed for a phase response that was precise to about2.5 rms, or about one part per million over the 5120cycles of the waveform. This response yielded near-in-range sidelobes in the 35-dB range, whereas far-outsidelobes rapidly fell to better than 40 to 50 dB down,as shown in Figure 4. In Figure 5, which is a photo-graph of a RAC developed for ALCOR, the two rain-bow-colored stripes near the centerline of the crystalshow light that is diffracted from the etched grating.The phase-compensating varying-width metal filmstrip runs down the centerline of the crystal.

Pairs of approximately one-inch-long matchedRAC devices were installed in ALCOR in 1974 and

were used successfully in a series of reentry tests.These devices proved to be such powerful wide-band-

width signal processors that advances in analog-to-digital converter technology to capture the output

were required before the capability of the RAC de-vices could be fully utilized.

RAC Pulse Compressors for the MASR Airborne RaFollowing the positive results with the early RAC de-vices, SAW technology was considered for a numberof Lincoln Laboratory programs. As the technology matured, the Laboratory SAW group helped guidethe development and procurement of SAW devicesfrom outside companies. Some device specificationsfell outside the state of the art, however, and so theinitial development of these more challenging devices

was carried out at the Laboratory. One example wasthe pulse compressors required for the experimentalMultiple-Antenna Surveillance Radar (MASR), anairborne radar for ground surveillance (see the articleentitled Displaced-Phase-Center Antenna Tech-nique, by Charles Edward Muehe and Melvin La-bitt, in this issue). This radar employed a 2.5-MHz-bandwidth pulsed linear-FM waveform with aduration of 125 sec.

The long pulse in the MASR proved to be a chal-lenge for SAW technology. A new material, bismuthgermanium oxide, with a low acoustic velocity wastried. A host of detailed technical obstacles were over-come in order to adapt the RAC technology to thisnew substrate material [24]. The package developedfor MASR incorporated three matched devices: apulse expander and two weighted pulse compressors.Phase errors were less than 2 rms, yielding better

FIGURE 5. The ALCOR RAC processor. The two rainbow-colored stripes near the centerline of this device are createdby the diffraction of light off the pair of etched gratings. Thevarying-width metal film strip running along the centerlineof the device performs phase compensation. This device re-placed the entire seven-rack processor shown in Figure 3.

Etchedgrating


6/24


7/24



micrometer technology group was set up at the Labo-ratory to pursue advanced lithographic techniques.

When interest in this area grew on the MIT campus,the Laboratorys expertise was called upon in the es-

tablishment of the Microsystems Center at MIT. Inaddition to transferring lithographic technology, theLaboratory has continued its own role of leadershipin microcircuit fabrication techniques.

Digital Signal Processing The development of digital signal processing for radarat Lincoln Laboratory provides a classic example of interdisciplinary technology transfer. The original ef-forts of researchers at Bell Telephone Laboratories,and by Bernard Gold and Charles Rader at Lincoln

Laboratory [33], were motivated by the desire tobandwidth-compress speech for more efficient digitalsecure-voice communication and to digitally simulateanalog components. This work led to Gold andRaders seminal book on digital signal processing[34]. The techniques developed during this time werevery powerful, and their immense applicability to sig-nal processing for ballistic missile defense becamereadily apparent [35].

The key realization of the potential for digital sig-nal processing in radars was the understanding thatballistic-missile-defense radars are pulsed systemsand, unlike analog signal processing, the digital signalprocessing did not need to be time synchronous. If raw data are digitized [36] and stored in memory, theavailable processing time is the time until the nextmeasurement, not the real-time extent of the mea-surement itself. This approach, then as it is now, is acareful balance among the required algorithms, an ar-chitecture that efficiently but flexibly implementsthose algorithms, and the selection of a hardwaretechnology that meets timeline requirements. Ex-amples of both the programmable and special-pur-pose approaches to radar signal processing are de-scribed below.

The Fast Digital Processor In the mid-1960s the emergent field of digital signalprocessing was becoming more well known. Excitingnew techniques for designing and implementing digi-tal filters were being published, and the fast-Fourier-

transform (FFT) algorithm in its various incarnationsoffered the prospect of drastically reducing the num-ber of computations necessary to perform importantsignal processing functions digitally (primarily multi-

plications, which were time-consuming operations ona general-purpose computer).

At Lincoln Laboratory there was growing frustra-tion among researchers over the inadequacy of thegeneral-purpose computer technology of the day forperforming digital-signal-processing calculations

with any kind of reasonable speed, notwithstandingcomputationally efficient algorithms such as the FFT.Thus in 1967, a team led by Gold, Rader, and PaulMcHugh conceived the architecture and instructionset for the Fast Digital Processor (FDP) [37, 38]. Al-

though, as mentioned above, a driving motivation was to simulate developmental speech-coding algo-rithms in real time or near real time, the overarchinggoal of the project was to achieve a design represent-ing an optimum balance for digital-signal-processingapplications between the computation throughput-rate potential offered by a purely special-purpose ar-chitecture and the flexibility afforded by a general-purpose computer. The result was a programmablemachine, architecturally optimized for digital-signal-processing computations, that offered the prospect of approximately two hundred times the throughputrate of a general-purpose computer for many digital-signal-processing applications through a combinationof advanced digital integrated-circuit technology (emitter-coupled logic), architectural parallelism, in-struction pipelining, and clever specialized architec-tural features (e.g., a bit-reversed add to facilitateradix-2 FFT address calculations).

The FDP architecture, illustrated in Figure 6, useddistinct structures for the program and data memo-ries, and it used a semimicrocoded instruction set.The FDP featured a 512 36-bit program memory to support the wide instruction-word format, which

was physically separate and distinct from two simul-taneously accessible 1024 18-bit data memories(extendable to 4096), all of which were implemented

with semiconductor-memory technology. The archi-tecture also incorporated four identical 18-bit, twos-complement, fixed-point arithmetic elements, as il-lustrated in Figure 7, which could be operated


8/24



concurrently and independently by virtue of the lati-tude provided by the 36-bit-wide instruction word.The FDP designers were among the first in the fieldof digital signal processing to recognize the so-called

multiply-accumulate operation as the most elementaldigital-signal-processing computational buildingblock, and the complex multiply as fundamental toFFT calculations. Therefore, the arithmetic elements

were configured and interconnected to facilitate thesecritical types of operations. The FDP was alsoequipped with flexible and powerful data-memory address-calculation mechanisms to further enhance

efficiency and performance for a wide class of digital-signal-processing functions.

The timing of the FDP was based on a three-deepinstruction pipeline comprising three 150-nanosec-ond epochs, which overlapped instruction fetch withinstruction decode/data-memory access and arith-metic-element operations. In principle, it was pos-sible to perform four arithmetic operations and fourlocal-data transfers per 150-nanosecond epoch, repre-senting a peak theoretical throughput rate of approxi-mately 53 million instructions per second (MIPS).

The four-quadrant multiplier, the single most costly component in the arithmetic elements in terms of hardware complexity, was implemented as fully in-stantiated combinatorial-logic arrays based on amodified Booths algorithm, and required 450 nano-seconds to produce a signed 36-bit product. To miti-gate this extra delay, other operations could be con-ducted within an arithmetic element while amultiplication was in process.

FIGURE 7. FDP arithmetic-element structure. The designshowed parallelism in several forms, including dual datamemories, four identical arithmetic elements, and a separateprogram memory. These features provided enhanced perfor-mance, particularly when computing complex arithmetic.

Datamemory

Output selection gates

18 18 array multiplier

Arithmetic/logic unit

Register Q N

MultiplierMultiplicand

Register IN

Register R N

Todata memory

N = 1, 2, 3, 4

R N + 1R N 1

C 18

R N 1R N + 1

IN 1 IN + 1

N + 1

C 18N

FIGURE 6. The Fast Digital Processor (FDP) architecture.The FDP itself comprised approximately 15,000 emitter-coupled-logic integrated circuits, dissipated about 2.5 kilo-watts of power, and occupied about 200 cubic feet of vol-ume. As technology evolved, an equivalent amount ofcomputing power could be realized in a few cubic feet. Suchmachines were known as array processors .

MD (right)MD (left)

MBMA

8channels

8channels

Bank switching

F register

X

12

12

1616

12

18

1024 1024

18

16

Instruction Register

MC (right)

Univac 1219

Channel 1(out)

Channel 0(out)

Channel 0(in)

Channel 1(in)

E0 register

0MC (left)256

18 1818180

MC (right)1MC (left)

1

256

XAU

Arithmeticelement 2

Arithmeticelement 1

Arithmeticelement 4

Arithmeticelement 3

MD (left) MD (right)

E1 register

Multiplexer Demultiplexer


9/24



Actual design and fabrication of the FDP were car-ried out at Lincoln Laboratory during the time framefrom 1968 to 1970, and represented no mean engi-neering feat. Some of the innovative layout and pack-aging concepts incorporated in the FDP came fromthe people in the Engineering division who had beenbuilding the Lincoln Experimental Satellites (LES).To achieve the desired performance goals for the FDP,the design and fabrication team needed to capitalizeon the then state-of-the-art Motorola MECL IIsmall-scale and 10k medium-scale digital integrated-circuit technologies. This effort required the develop-ment of novel and sophisticated design methodolo-gies heretofore unheard of in digital systemimplementations, because of the high speed of thelogic and the finite speed of electrical-signal propaga-tion. For example, all data, control-signal, and clock-distribution paths required careful attention to physi-cal length, signal quality, and impedance control for

reliable and predictable operation. The design prac-tices pioneered in the construction of the FDP even-tually became commonplace within the digital designcommunity as experience with ultrahigh-speed digi-tal-circuit technology grew. Figure 8 shows the fin-ished FDP facility, which included a Univac 1219general-purpose host computer. The FDP propercomprised approximately 15,000 integrated circuits,dissipated about 2.5 kilowatts of power, and occupiednominally 200 cubic feet of space.

Although not easy to program, the FDP proved tobe a unique, versatile, and powerful asset, as had beenhoped. For example, a two-pole digital resonator or aradix 2 FFT butterfly could be executed in approxi-mately 1.2 sec. The architecture, though optimizedfor digital filtering and FFT computations, was stillgeneral enough to be useful for other types of nu-meric computation, and it even supported extended-precision and floating-point operations. As a testi-

FIGURE 8. The FDP facility at Lincoln Laboratory in 1970, which included a Univac 1219 general-purpose hostcomputer. The arithmetic/logic unit incorporated a full 18-bit, twos-complement adder/subtractor, supported allBoolean functions, and included linkages for extended-precision calculations. The 18 18-bit four-quadrant mul-tiplier was based on a modified Booth s algorithm, and was implemented as a full combinatorial array usingsingle-bit adders.


10/24



mony to its flexibility, the first real-time implementa-tion of a 2400-bit-per-second linear predictive speechcoder (LPC), which involved numerical computa-tions far less regular and structured than those of adigital filter or an FFT, was successfully demonstratedon the FDP in the early 1970s [39, 40].

This work led to a series of increasingly compactspecialized digital signal processors for real-timeimplementation of LPC and other digital voice-com-pression algorithms, such as the first stand-alone LPCvocoder shown in Figure 9. This work culminated inthe DARPA-sponsored Speech Processing Peripheral,

which was a direct precursor to the next generation of secure telephone units, or STU-IIIs, introduced intogovernment service during the early 1980s.

The FDP also proved useful in radar signal pro-cessing applications, where it was capable of real-timeperformance if appropriate specialized adjunct hard-

ware components were provided when necessary (e.g., an external corner-turning buffer memory) andthe range-Doppler space of experimental interest wassuitably restricted. For example, in the early 1970sthe Federal Aviation Administration (FAA) was ex-ploring signal processing techniques that might pro-vide a cost-effective performance upgrade for the then

existing generation of airport surveillance radars(ASR). The FDP was connected to a remote ASR transceiver through a custom-designed duplex datalink, and was used for the development and real-time

evaluation of novel Doppler-processing techniquesfor clutter mapping. The FDP simulation experi-ments proved that a special-purpose digital-signal-processing hardware adjunct to the ASR sensor couldbe both effective and economical [41, 7]. Also, in asimilar time frame, the FDP/data-link facility wasused as part of the Long-Range Demonstration Radarproject to develop moving-target-indication (MTI)algorithms for surface vehicles or other relatively slow-moving objects amidst heavy ground clutter indefense applications. In particular, these processors

allowed engineers to implement and performreal-time evaluations of experimental Doppler-pro-cessing, post-detection integration, and statistical-de-cision algorithms [42].

Although it was a one-of-a-kind machine, the FDPproved the value of programmable machines orientedtoward digital signal processing, and it served as amotivator for the first generation of commercial off-the-shelf programmable digital-signal-processing ac-celerators that reached the marketplace during the1970s, offered by such manufacturers as ComputerSignal Processors, Signal Processing Systems, andFloating Point Systems.

Digital Convolver System An early application of the special-purpose approachto digital signal processing arose from initial researchfor the U.S. Army in the early 1970s on an all-solid-state radar for ballistic missile defense, which led tothe development of a conceptual L-band radar calledthe Advanced Fielded Array Radar (AFAR). The L-band radar concept used solid state transmit modules;consequently, it required long waveforms for detec-tion as well as short waveforms for tracking. The needfor a large bandwidth to provide adequate range reso-lution led to a large waveform repertoire with a widediversity of time-bandwidth products. This repertoireprecluded the use of analog filters, in that there sim-ply would have been too many fixed filters.

The flexibility of digital signal processing [43] sug-gested that a suitable digital processor design would

FIGURE 9. The first stand-alone compact linear predictivespeech coder, or LPC vocoder, which served as a majordriver and motivating force for the next-generation commer-cial secure telephone units (STU-III) introduced into gov-ernment service during the early 1980s. This vocoder wasbased on the state-of-the-art commercial single-chip digital-signal-processing microcomputers available at that time.


11/24



provide a solution, as long as the processing could bedone in real time (i.e., in the total time available). Theresult was the Digital Convolver System (DCS) [44],

which was intended to provide the required flexible

real-time matched filtering of large numbers of wave-forms, some with large time-bandwidth products.The design was based on fast-convolution techniques[45], and provided for a 16,384-point radix-4 FFT,clocked at 30 MHz, to achieve a throughput data rateof one 16k FFT every 136 microseconds [46].

Two innovations at the time were the use of a hy-brid floating-point data format and CORDIC (coor-dinate-rotation digital computer) [47] rotators in theFFT. The hybrid floating-point format uses a com-mon exponent for both the real and the imaginary

parts of the complex data at each stage of the FFT cal-culation, and it was sometimes referred to as vectorfloating point. This approach greatly alleviated thecomputational hardware complexity of the system[48, 49]. Similarly, the CORDIC rotator provided acomputationally efficient implementation of thecomplex multiplications required in the FFT. An-other innovation was based on the observation that

FIGURE 10. The Digital Convolver System (DCS) architecture. This system exploits the fact thatDoppler processing of radar waveforms uses Doppler-shifted versions of a single reference func-tion. Consequently, if the processing is performed by fast convolution, only one forward transform isneeded. The result is stored, read multiple times, Doppler-shifted, and inverse-transformed multipletimes. The forward and inverse transforms are both performed in the reconfigurable pipeline fast-Fourier-transform (FFT) subsystem shown in the figure.

Reference-functionmemory

Coefficientmemory

Interstagedelay

memoriesand

switches

Stage 1 2 3 5 6 7

A/D

60 Ms/sec10 bits

120 Mwds/sec3.24 Gbits/sec

In v

Temporary storagememory (16k)

Out

Reconfigurable pipeline FFT

4

Inputbuffer

Four-pointdiscreteFourier

transform

e j

ve j

Doppler processing using the fast-convolution ap-proach did not require the repetitive use of the for-

ward FFT. Rather, a single forward transform fol-lowed by multiple inverse transforms was sufficient.

The resulting reduction of the hardware requirements(by roughly one half) was significant.

Figure 10 illustrates the DCS architecture. Thesystem includes a temporary storage memory, a refer-ence-function memory, and a multiplier system. Thetemporary storage memory holds the forward-trans-formed data and sends the data through the fre-quency-domain multiplier for multiple inverse trans-forms. The core of the system is the pipelined FFT[50, 51], which is shown in detail in Figure 11. Themost important feature of this system is that the

interstage delay-line memories are reconfigurable, which allows the same set of hardware to provideboth forward and inverse transforms of 4k, 8k, or 16k points, while also allowing the data to be read into theforward FFT and out of the inverse FFT in normalorder. Figure 11 shows seven elementary computa-tion elements and six interstage-delay memory ele-ments, which are reconfigured depending on the size


12/24



of the transform and whether a forward or inversetransform is being performed. This process is indi-cated by the two major paths through the figure.

The concept of implementing the signal process-ing by using digital technology was relatively new atthe time. The potential for achieving highly accurateprocessing, however, was enormous. The DCS dem-onstrated and certified this potential by achieving acomputational noise floor with spurious peaks ap-

proximately 63 dB down, as shown in Figure 12, a re-sult that proved the viability of the hybrid floating-point approach.

The DCS [52] used mostly emitter-coupled logic10k-series integrated circuits to meet the throughput-rate requirements. One large multiplexed memory,however, used MOS technology, and there were a few transistor-transistor-logic interface circuits. The DCShad about 27,500 integrated circuits and consumed

FIGURE 11. The reconfigurable DCS FFT architecture. This system is designed to allow the same hardwaresubsystems to perform multiple transform sizes (4k, 8k, and 16k) and simultaneously perform both the forwardand inverse transforms. The penalty is an increased amount of data routing, but this penalty is more than out-weighed by the savings in hardware that would be incurred if two complete transform systems had to be built.

256, 512, 1k

1k

1k, 2k. 4k , 2k, 4k

64, 128, 256256

Interstagedelay memories

Coefficientmemory

Computationelements

1, 2, 44

3072 48

12k, 4k 12

768 192

1 6 k F

1

4 k F 1

Al l F

4 k F

A

l l F 1

1 6 k F

8 k F

8 k F 1

4, 8, 1616

16, 32, 6464

Inverse-transform input

Forward-transform

inputs

Inverse-transform

outputs

Forward-transform output

Coefficientmemory

Coefficientmemory

Coefficientmemory

Coefficientmemory

Coefficientmemory


13/24



FIGURE 12. The DCS computational noise floor is achievedby using hybrid floating-point arithmetic. The resultsachieved by the DCS demonstrated that digital-signal-pro-cessing techniques have a performance potential limitedonly by the word length used.

FIGURE 13. The DCS in 1979. At that time it was the fastest and largest pipelined FFT proces-sor ever built. The system was large; it was comparable to the ALCOR all-range processorshown in Figure 3, but with an order-of-magnitude improvement in performance.

90

100

80

70

60

50

40

30

20

10

0

10

R e

l a t i v e p o w e r

( d B )

Number of samples

8192

Waveform type:linear frequencymodulation

Bandwidth = 10 MHz

71686144512040963072204810240

Time-bandwidthproduct = 4096

Weighting: Blackman

15 kW of power. At its completion in 1979, the DCS,shown in Figure 13, was the fastest and largest pipe-lined FFT processor that had yet been built.

The FAA: The Moving-Target Detector and theParallel Microprogrammed ProcessorIn 1972 the FAA brought a radar problem to LincolnLaboratory. The FAA was in the process of developingthe Automated Radar Terminal System (ARTS-3),

with the aim of computerizing air-traffic-control dis-plays at airports. They had successfully automated the

Air Traffic Control Radar Beacon System, and in sodoing provided automatic track acquisition and up-dating for all beacon-equipped aircraft (secondary radar). They had been unsuccessful, however, in auto-

mating the primary, or skin-tracking, radar. The pri-mary radar produced too many clutter-related falsealarms and missed detections as a result of the tech-niques employed to deal with the clutter.

With the advent of medium-scale integrated cir-cuits around 1970, many new signal processing algo-


14/24



rithms were developed. This evolving integrated-cir-cuit technology allowed digital sampling and filteringof an ASRs single-scan output in over three millionrange-azimuth-Doppler cells. Thresholding algo-

rithms (which are described later in this article) couldthen be employed for the type of clutter found ineach resolution cell (i.e., ground clutter in each zero-velocity Doppler cell could be thresholded by using adigitally stored ground-clutter map), thus avoidingfalse alarms while keeping all of the resolution cells assensitive as possible for the detection of aircraft. Thistype of processor was named the moving-target detec-tor (MTD) to distinguish it from the now old-fash-ioned moving-target indicator (MTI). An initial exer-cise using Lincoln Laboratorys FDP [37, 41] verified

the usefulness of these algorithms over a small eight-nautical-mile by 45 sector. This advance was fol-lowed by full-scale development and testing of theMTD, led by Charles Edward Muehe. Two versionsof this processor were built. In the MTD-1 the algo-rithms were hard wired into the processor [53], andin the MTD-2 [54] the algorithms were implementedas software in a parallel microprogrammed processor(PMP) [55]. The MTD-2 found its way into at leastsix different types of surveillance radars, includingboth ground-based and airborne radars.

The MTD Class of Radars The MTD radars incorporate a number of novel sig-nal processing techniques. The older MTI radarsstaggered pulse-repetition-frequency (PRF) wave-form, which was used to ameliorate blind speeds, isreplaced in both the MTD-1 and the MTD-2 by amultiple-PRF waveform, wherein about eight pulsesat one PRF in a coherent processing interval are alter-nated with a coherent processing interval with a 20%different PRF. The receiver maintains linearity overthe full dynamic range of the analog-to-digital con-verters. For each coherent processing interval a bank of digital filters, each designed to maximize the sig-nal-to-clutter ratio, is implemented in each rangegate. Several forms of detection thresholding are used,depending on the statistics of the expected clutter re-flections in each filter. An algorithm is employed toflag range gates that contain interfering pulses.

To cause a uniform presentation on the plan-posi-

tion-indicator (PPI) display in the presence of groundclutter, older MTI radars employed amplifiers in theMTI channel that were limited to about 20 dB abovethe receiver noise [56]. This limiting spreads the clut-

ter spectrum and reduces the MTI subclutter visibil-ity to at most about 20 dB. The MTD, on the otherhand, has a measured subclutter visibility of 42 dB,

which is in turn limited by the receivers dynamicrange. Because the spatial statistics of ground clutterare highly non-Gaussian, both MTD radars use aclutter map for thresholding the zero-velocity Dop-pler filter. Older MTI radars have a notch-at-zeroDoppler, and thus they cannot detect a crossing targetthat has a near-zero radial velocity. The clutter mapallows detection of crossing aircraft, which would

usually present large reflections from their fuselages when crossing or are in a low ground-clutter regionbecause of ground shadowing. As a consequence of this detection capability, the MTD is said to havesuperclutter and interclutter visibility.

The high pulse-to-pulse correlation of rain-clutterreturns, together with noncoherent binary integra-tion, caused the sliding-window detector used inolder MTI radars to exhibit a high false-alarm rate inrain. The strictly coherent integration for each of theMTDs nonzero Doppler filters, together with thresh-olds based on the mean clutter level within 0.5 nmiof each thresholded range-Doppler cell, keeps theMTDs false-alarm rate under excellent control. Theupdate of the zero-velocity ground-clutter thresh-olding map is adjusted so that it also keeps up withchanging rainstorm backscatter as the storm passesthrough the radars coverage. Because multiple PRFsare used, the target appears in a different filter on suc-cessive coherent processing intervals (unless it has thesame radial velocity as the storm), resulting in a goodchance of detection. The MTDs constant PRF ineach coherent processing interval, instead of the olderMTI radars staggered pulse-repetition intervals, al-lows the illumination of second-time-around clutter,

which is filtered in the same way as close-in clutter.For each threshold crossing, a primitive report is

sent to the MTDs post-processor, giving the ampli-tude, range, azimuth, Doppler-filter number, andPRF. Reports that appear to come from the same tar-get are interpolated for the best estimate of the targets


15/24



amplitude and position and are used for target-track initiation and updating. Also in the post-processor,area thresholds are maintained to control excess falsealarms, particularly from bird flocks. Each area of about sixteen square nautical miles is divided intoseveral velocity regions. The threshold in each regionis adjusted on each scan to achieve the desired limiton false alarms without raising the threshold so highthat small aircraft are prevented from being placed intrack status. The post-processor also implements amap of small areas, only a few resolution cells in ex-tent, in which the clutter return is so high that falsealarms occur repeatedly. Detection in these areas iscensored.

The MTD-1 was initially tested at Lincoln Labo-ratory by using an S-band AN/FPS-18 radar with aklystron transmitter that had been modified to im-prove its stability. The MTD-1, which is shown in

Figure 14, was transferred in late 1974 to the FAAsradar test facility near Atlantic City, New Jersey,

where it was connected to an ARTS-3 radar. The FAA [57] and Lincoln Laboratory engineers tested theMTD-1 extensively. Figure 15 shows the results of ra-dar detection tests of small aircraft in rain. Figure15(a) shows the extent of rain clutter, and Figure15(b) shows the detection and automatic tracking of a number of aircraft for about four minutes. The ver-tical track at the center is detection of automobiles ona road. Later improvements included automatic

elimination of moving road vehicles. A competition[58] was held between the MTD and the RVD-4,

which was an advanced version of the sliding-window detector that estimated the correlation of rain-clutterreturns and readjusted the thresholds appropriately.In this competition the MTD radars false-alarm andtarget-detection performances proved to be markedly superior to those of the RVD-4.

In December 1975, the U.S. Air Force Air DefenseCommand arranged to test the MTD-1 in the pres-ence of active electronic countermeasures and chaff [59]. An Air Force EB-57 equipped with four hun-dred pounds of chaff along with swept, spot, and bar-rage jammers was used for the test. The EB-57 andanother test aircraft were detected with nearly unity blip scan ratio as they flew through the chaff. Thesetests demonstrated the superior detection perfor-mance of the MTD-1 in chaff and jamming, accom-panied by a low false-alarm rate.

With the establishment of the superior perfor-mance of these techniques in both military and civil-ian environments, it was not long before contractors

were proposing using these techniques on most new air-defense radars and on new developments in air-traffic-control radars.

The MTD-2 and the Parallel Microprogrammable Processor By 1975 the FAA had decided that the MTD class of radars was an effective solution to the problem of de-tecting aircraft in high-clutter environments, but

FIGURE 14. The moving-target detector (MTD-1) at the Fed-eral Aviation Administration (FAA) facility in Atlantic City,New Jersey, in 1974. The MTD-1 was extensively tested incompetition with a modern digitized version of the moving-target indicator (MTI) delay-line canceler.


16/24



there were reservations concerning its complexity. Be-cause the algorithms were embedded in the hardware,it would take a digital engineer or a highly trained ra-dar technician to diagnose troubles. Lincoln Labora-tory was encouraged to consider alternative designsthat would relieve the logistic and maintenance prob-lems that might arise. At that time, the concept of parallel processing was just evolving, and the notionthat many signal processing problems lent themselvesto architectures that applied a single, relatively rudi-mentary algorithm to multiple data sets was one of the innovative realizations of the power of digital sig-nal processing. The parallel microprogrammed pro-cessor, or PMP [55, 60], was an important early ex-ample of this kind of architecture.

The PMP was an SIMD (single-instruction mul-tiple-data-stream) computer consisting of a numberof processing modules (typically two to eight), allserved by one control unit. This type of system wasseen as particularly appropriate for a surveillance ra-dar such as the ASR, because the same algorithms areused for each range gate. One PMP module servedten nautical miles of range in an ASR. An extra pro-cessing module served as a spare, to be switched in

when a fault was detected in the primary module. A processing module consisted of two wire-

wrapped boards: one to hold the input data and clut-ter-map memories; the other, the processing element,to handle all the mathematical computations. Theprocessing element contained two 24-bit arithmeticand logic units, a bit shifter, and a small high-speedmemory. The processing element operated with a 75-nsec instruction cycle, and on average it performedtwo simple operations per cycle time, resulting in anet processing rate of 25 million instructions per sec-ond. The control unit also consisted of two wire-

wrapped boards. One board held memory for instruc-tions, program constants, and target reports from theprocessing modules. Its processing element did all therequired arithmetic, such as memory-address genera-tion and time keeping, and interfaced with the pro-cessing modules and the post-processor. To handlethis kind of computational workload, a PMP assem-bly language was developed at Lincoln Laboratory.Each line of code contained all the assembly languageinstructions to be executed in one cycle time. Themachine language was generated by using a cross-compiler that was also written at Lincoln Laboratory

FIGURE 15. Performance of the MTD in heavy precipitation and ground clutter. This figure shows the detection of a small targetaircraft in rain (a) with normal video, before the installation of the MTD, and (b) after the installation of the MTD. Notice the ab-sence of false returns and the continuous tracking in the MTD image, even of aircraft with zero radial velocity. The target air-craft is a single-engine Piper Cherokee.

(a) (b)

Small targetaircraft


17/24



and executed on the Laboratorys central computer.Three PMP-1 devices were built at Lincoln Labora-tory and seven PMP-2 devices, as shown in Figure 16,

were built under contract by Stein Associates.

In 1978 Lincoln Laboratory installed an MTD-2,using a PMP-2 with the ASR-7 radar, at theBurlington, Vermont, airport [61]. The FAA chosethis site near Mount Mansfield because it is reputedto have the worst clutter environment on the eastcoast. The FAA brought in air traffic controllers andother personnel from all over the United States to ob-serve and operate the Laboratorys MTD-2 radar.Convinced that the MTD-2 was what they wanted,the FAA asked Lincoln Laboratory to help write thespecifications for the next-generation ASR. A produc-

tion contract was placed with Westinghouse for the ASR-9 radar, which today is in operation throughoutthe United States. The development of the MTDconcept fundamentally changed the way surveillanceradars are designed, and it caused that change quickly,essentially overnight! As a result of its successfulimplementation, the acronym MTD has since be-come an eponym.

Space-Based Surveillance of the EarthDuring the 1970s the Communications division atLincoln Laboratory examined ways to reduce the vul-nerability of military communication satellites to

jamming. This need led to the study of adaptive-null-ing techniques to minimize the effect of jamming. In1985 the Laboratory began studying the feasibility of a large array radar that would search for movingground or airborne targets from low-earth orbit. Thisproposed orbiting-radar design is another example of interdisciplinary technology transfer, showing how the expertise developed from the communications-satellite effort could be applied to problems in radarsignal processing

A space-based surveillance radar must handle twomajor sources of interference: clutter from the entirevisible earth and jamming in the antenna sidelobes.The clutter can be attenuated by using displaced-phase-center-antenna techniques (see the article by Muehe and Labitt in this issue); the sidelobe jammerscan be attenuated by modifying some of the array-ele-ment weights to cause deep pattern nulls in the jam-

ming directions. In both cases the formation of asingle data stream from current and delayed datafrom many antenna elements requires the computa-tion of a weighted sum for each sampling instant.This computation can require a very large number of multiplications and additions per second: four timesthe product of the number of antenna elements andthe sampling rate. Because these operations are regu-lar, it was determined that they could be imple-mented by using commercially available special-pur-pose integrated circuits.

The determination of the appropriate set of weights is another matter. These weights must be de-termined adaptively. As the satellite moves relative tothe surface of the earth, each jammer appears to move

FIGURE 16. The parallel microprogrammed processor(PMP) built by Stein Associates and installed at the Burling-ton, Vermont, airport in 1978. This detector was displayed tovisiting air traffic controllers from all over the United States,who were positively impressed with its performance.


18/24



from one part of the sidelobe region to another, andtherefore the weights must be readapted about twohundred times per second. An algorithm to computethese adapted weights is much more complicated

than a simple sum of products, and in 1985 it seemedto require a computer capable of adding, subtracting,multiplying, dividing, computing square roots, andstoring large amounts of data. At that time single-chip digital-signal-processing computers were avail-able, but they were many times less efficient than thesimple special-purpose chips for computing sums of products. The cost of carrying out the weight-adapta-tion algorithm depends sensitively onN , the numberof weights being determined. The computational costis proportional to the cube of N , so that determiningthe weights for twice as many antenna elements re-quires eight times the number of multiplications andadditions.

Lincoln Laboratory engineers working on thisproblem in 1985 therefore estimated that it would bereasonable to fly enough computing power to adapttwenty-five weights, though there were many reasons

why system designers might have wanted to use alarger number. In a very narrowband system withmodest aperture, for example,N + 1 weights are re-quired to null out N jammers. If the bandwidth of theradar is larger, or if the array aperture is large, several

weights can be required per nulled jammer. Adapta-tion to clutter also requires many weights.

In the same year a small project was initiated todevise and demonstrate an efficient approach to thecomputation of adaptive weights. The result was thediscovery, early in 1986, of an unique confluence of atechnology, an algorithm, and an architecture thatenabled the construction of an adaptive weightingcomputer called MUSE (Matrix Update Systolic Ex-periment). MUSE, a demonstration system, was ca-pable of computing sixty-four weights several hun-dred times per second, but it had a physical size and

weight no larger than a package of cigarettes. At thattime, no actual adaptive antenna arrays with sixty-four elements existed: it would have made no sense tobuild such arrays, since nothing (i.e., no existingcomputer) could adapt their weights in real time.

The data used to determine the weights in theMUSE algorithm are a series of columns of complex

numbers. Each column contains one sample fromeach of the N antenna elements. It is important tounderstand that the data arrive one datum at a time,one column at a time. This limited serial data transfer

means that the number of data input pins required isquite reasonable.

The computation of the adaptive weights involvesthe triangularization of the raw data and a back-sub-stitution that yields the actual weights. The triangu-larization process is in essence a sequence of two-di-mensional rotations. These rotations are appliedsequentially to the original data matrix until the ma-trix has all zeros in the upper-right portion and nozero values in the lower-left portion. The solution of the weights using back-substitution is then algorith-

mically straightforward and computationally simple.Given that the critical part of the adaptive-weight

computation can be reduced to a sequence of simplerotations, it became important to look for efficient

ways to implement such a rotation. A design for sucha rotating circuit was developed in the 1950s, and it iscalled a CORDIC module [47]. The CORDIC mod-ule is made up of adders and shifters, and it is easily pipelined so that it can accept new pairs of numbersas fast as it can add, even though any rotation takesmuch longer than any addition. A CORDIC moduleis a convenient size to be realized as a single integratedmodule. All ninety-six CORDIC modules requiredfor MUSE are identical and can be easily intercon-nected. In this way the architecture of MUSE and thealgorithm it carries out are perfectly adapted to eachother.

A further improvement was the use of wafer-scaleintegration. This technology had been attempted by many laboratories in the 1980s, but Lincoln Labora-tory was the first to succeed in building wafer-scalecircuits [62]. The difficulty with wafer-scale integra-tion is that even one tiny defect on a chip usually makes the chip nonfunctional. When the chip is a

whole wafer, the probability of a defect becomes a vir-tual certainty. The Laboratorys approach was to builda wafer with redundant cells and to connect togetherenough of each type of cell to yield a working system.In the case of MUSE, there was only one type of cell,a CORDIC module. A wafer was fabricated with 132CORDIC modules. Interconnections were made by


19/24



FIGURE 17. The MUSE (Matrix Update Systolic Experiment) wafer provided an efficientapproach to the computation of adaptive weights. This demonstration system could com-pute sixty-four adaptive weights several hundred times per second.

using an automated laser weld to make electrical con-nections between the cells. The same automated laser

was used to break connections, when necessary, by vaporizing metallization.

The active area of a MUSE wafer, shown in Figure17, fit into a square of just over three inches on a side(or nine square inches in area). At a clock rate of 6MHz, the system was able to carry out almost threehundred million rotations per second, equivalent toabout three billion instructions per second in a con-ventional single-instruction computer. Power con-sumption was only about 10 W, and because there

were so few wired connections, MUSE was a highly reliable design suitable for space applications.Through further refinement of the integrated-circuitfabrication technology, a modern version of MUSE

developed by the Hughes Corporation has one thou-sand times the computational power of Lincoln Lab-oratorys original demonstration.

Summary The proliferation of radar signal processing efforts atLincoln Laboratory has been driven by the over-

whelmingly dominant need to detect and measurefundamentally small radar-target returns in the pres-ence of potentially overwhelming noise and other un-

wanted returns (i.e., clutter, both natural and inten-tional). This requirement has fundamentally involvedthe concurrent development of (1) theory and algo-rithms, (2) the underlying analog and digital technol-ogy [63], and (3) efficient architectures that mergetheory and device technology into real systems for


20/24



important military and commercial applicationsonthe ground, in the air, and in space. These develop-ments, which started with what might now be viewedas primitive efforts in SAGE and early ballistic missile

defense, progressed through the development of fun-damental device technology, both analog and digital,and have now moved in the direction of exploitingthe enormous power and flexibility of digital process-ing, both custom and commercial.

For example, efforts are under way to develop ex-tremely high-performance systems that combine clas-sical clutter suppression with computationally chal-lenging adaptive processing for joint detection of targets in clutter and jamming (a technique known asspace-time adaptive processing, or STAP) [64]. More-

over, recent successes in radar imaging hold the prom-ise for real-time and near-real-time generation of complex images that could be exploited by analystsfor rapid adaptation to evolving circumstances. Thesecombined techniques doubtlessly will find their way into future advanced ground, airborne, and spacesystems.

In viewing the history of signal processing, we notean interesting paragraph in Merrill Skolniks 1962seminal book on radar [65]: The maximum com-pression ratios possible will depend upon the amountof development effort expended to achieve them. Thenumerical examples given by Krnert [66] for Gaus-sian-shaped pulses and cascaded-lattice networks in-dicate the feasibility of achieving pulse-compressionratios from = 8 to 40. In Darlingtons patent [67]an example is given for a Gaussian-shaped pulse in

which a compression ratio of 34 is mentioned. TheBritish patent issued to Sproule and Hughes [68]claims that it is possible to achieve a pulse-compres-sion ratio of 100. Klauder [3] et al. also suggest thatpulse-compression ratios of approximately 100 arepossible. The extraordinary advances in radar signalprocessing in the past five decades admit technology that today allow radars with significantly in excessof 1,000,000.

AcknowledgementsThe efforts described in this article span fifty years of research at Lincoln Laboratory. As such it is impos-sible to acknowledge the efforts of all the people at

the Laboratory, in government, and in industry whocontributed to, supported, and encouraged these pro-grams. The authors can only acknowledge the generalsupport of the U.S. Army, the U.S. Air Force, U.S.

Navy, DARPA (formerly ARPA), the Ballistic MissileDefense Organization (BMDO), and the FAA. How-ever, James Carlson, now retired from the BMDO,deserves special mention for his long-term keen inter-est and broad support of radar signal processing, notonly at Lincoln Laboratory, but across the country.

Within the Laboratory, Irwin Lebow contributed en-thusiasm and strength of leadership, and Ben Goldprovided intellectual guidance in the field of digitalsignal processing at a time when it was just emerging.Lastly, the lead author would like to remember the

late Jerry Margolin for his phenomenal insight andthe incredibly spirited discussions that he provided anew young staff member at Lincoln Laboratory.


21/24


22/24



51. G.C. OLeary, Nonrecursive Digital Filtering Using CascadeFast Fourier Transforms,IEEE Trans. Audio Electroacoust. 18(2), 1970, pp. 177183.

52. The DCS was designed by Lincoln Laboratory and manufac-tured by General Electric Heavy Military Systems Group to a

detailed design specification.53. W.H. Drury, Improved MTI Radar Signal Processor,LincolnLaboratory Project Report ATC-39 (3 Apr. 1975), FAA-RD-74-185, DTIC #ADA-010478/6.

54. L. Cartledge and R.M. ODonnell, Description and Perfor-mance Evaluation of the Moving Target Detector, LincolnLaboratory Project Report ATC 69 (3 Aug. 1977), DTIC#ADA-040055.

55. W.H. Drury, B.G. Laird, C.E. Muehe, and P.G. McHugh,The Parallel Microprogrammed Processor (PMP),Radar 77, London, 2528 Oct. 1977.

56. W.W. Schrader and V. Gregers-Hansen, MTI Radar, chap.15, Radar Handbook,M.I. Skolnik, ed., 2nd edition (McGraw Hill, New York, 1990), pp. 15.115.72.

57. R.S. Bassford, W. Goodchild, and A. DeLaMarche, Test andEvaluation of the Moving Target Detector, Final Report, Oct.1977, FAA-RD-77-118, DTIC #ADA-047887.

58. R.M. ODonnell and L. Cartledge, Comparison of the Per-formance of the Moving Target Detector and the Radar VideoDigitizer,Project Report ATC-70,Lincoln Laboratory (26 Apr.1977), NTIS No. ADA-040472.

59. R.M. ODonnell and L. Cartledge, Evaluation of the Perfor-mance of the Moving Target Detector (MTD) in ECM andChaff,Technical Note 1976-17,Lincoln Laboratory (25 Mar.1976).

60. G.P. Dinneen and F.C. Frick, Electronics and National De-fense: A Case Study,Science 195 (4283), 1977, pp. 11511155.

61. D. Karp and J.R. Anderson, Moving Target Detector (ModII) Summary Report,Project Report ATC-95,Lincoln Labora-tory (3 Nov. 1981), DTIC #ADA-114709.

62. C.M. Rader, Wafer-Scale Integration of a Large Systolic Array for Adaptive Nulling,Linc. Lab. J. 4 (1), 1991, pp. 330.63. An interesting snapshot of the state of the art of both analog

and digital processing technology, circa 1977, is contained inchap. 10, by R.J. Purdy, of Radar Technology,E. Brookner, ed.(Artech House, Dedham, Mass., 1977), pp. 155162.

64. J. Ward, Space-Time Adaptive Processing for Airborne Ra-dar, Lincoln Laboratory Technical Report 1015,Dec. 1994,DTIC #ADA-293032.

65. M.I. Skolnik, Introduction to Radar Systems (McGraw-Hill,New York, 1962), p. 495.

66. R. Krnert, Impulsverdicktung [Pulse Compression], pt. 1,Nachr. Tech. Elektron. 7, Apr. 1957, pp. 148152, 162; pt. 2,Nachr. Tech. Elektron. 7, July 1957, pp. 305308. For Englishabstractions of these two articles, see abstract 72,Proc. IRE 46

(2),1958; abstract 1078, Proc. IRE 46 (5), 1958, p. 936.67. S. Darlington, Pulse Transmission, U.S. Patent No.2,678,997, 18 May 1954.

68. Improvements in and Relating to Systems Operating by Means of Wave Trains, British Patent Specification 604,429,5 July 1948, issued to Henry Hughes and Sons, Ltd., D.O.Sproule, and A.J. Hughes.

35. R.J. Purdy, P.E. Blankenship, A.E. Filip, J.M. Frankovich, A.H. Huntoon, J.H. McClellan, J.L. Mitchell, and V.J. Sfer-rino, Digital Signal Processor Designs for Radar Applica-tions, Technical Note 1974-58,Lincoln Laboratory (31 Dec.1974).

36. The digitization process is crucial to the implementation of theprocessing. At the time, analog-to-digital (A/D) converters didnot exist with the requisite word length (dynamic range) andsample rate. Consequently, Lincoln Laboratory initiated a de-velopment with Hughes Aircraft for a 10-bit, 60-Msec/sec

A/D converter. This effort was highly successful. Four A/Dpairs (I and Q) were built. Eventually, one pair was used on the

Armys Signature Measurements Radar and two were used formany years on the Cobra Judy Radar. (See the article entitledRadars for Ballistic Missile Defense Research, by Philip A.Ingwersen and William Z. Lemnios, in this issue.)

37. B. Gold, I.L. Lebow, P.G. McHugh, and C.M. Rader, TheFDP, a Fast Programmable Signal Processor, IEEE Trans.Comput . 20 (1), 1971, pp. 3338.

38. L. Rabiner and B. Gold, Theory and Application of Digital Sig-nal Processing (Prentice-Hall, Englewood Cliffs, N.J., 1975).

39. E.M. Hofstetter, personal communication, Lincoln Labora-tory, ca. April 1974.

40. J.A. Feldman, E.M. Hofstetter, and M.L. Malpass, A Com-pact, Flexible LPC Vocoder Based on a Commercial SignalProcessor Microcomputer,IEEE J. Solid-State Circuits 18 (1),1983, pp. 49.

41. C.E. Muehe, Jr., L. Cartledge, W.H. Drury, E.M. Hofstetter,M. Labitt, P.B. McCorison, and V.J. Sferrino, New Tech-niques Applied to Air-Traffic Control Radars,Proc. IEEE 62(6), 1974, pp. 716723.

42. B. Gold and C.E. Muehe, Digital Signal Processing forRange-Gated Pulse Doppler Radars, XIXth AGARD Conf.Proc. on Advanced Radar Systems, No. 66, 2529 May 1970,Istanbul, Turkey.

43. One key advantage of digital processing is the ability to pre-cisely simulate the computations in advance. This approach was used extensively in the design of the Digital ConvolverSystem.

44. A.H. Anderson, J.M. Frankovich, L. Henshaw, R.J. Purdy, andO.C. Wheeler, The Digital Convolver System,Project Report SDP-228,Lincoln Laboratory (19 June 1981).

45. P.E. Blankenship and E.M. Hofstetter, Digital Pulse Com-pression via Fast Convolution,IEEE Trans. Acoust. Speech Sig-nal Process . 23 (2) 1975, pp. 189201.

46. B. Gold and T. Bially, Parallelism in Fast Fourier TransformHardware, IEEE Trans. Antennas Propag. 21 (1), 1973, pp.516.

47. J.E. Volder, The CORDIC Trigonometric Computing Tech-nique, IRE Trans. Electron. Comput . 8 (3), pp. 330334.

48. As part of an Air Forcesupported internal Lincoln Laboratory

effort, the Laboratory developed and fabricated a set of very fast2-bit adder/subtractor circuits that was ideally suited for fastsigned arithmetic array multiplication. This design was modi-fied by Peter E. Blankenship into a single programmable adder/subtractor component suitable for FFT and CORDIC rotatorcomputations. This design was subsequently transferred by theLaboratory to Motorola and incorporated in their MECL10K product line as the MC10287L.

49. S.D. Pezaris, A 40-ns 17-Bit by 17-Bit Array Multiplier,IEEE Trans. Comput. 20 (4), 1971, pp. 442447.

50. H.L. Groginski and G.A. Works, A Pipeline Fast FourierTransform, EASCON Record, Washington, 2729 Oct. 1969 ,pp. 2229.


23/24



.

received an S.B. degree fromMIT, and M.S. and Ph.D.degrees from Purdue Univer-sity, all in electrical engineer-ing. In 1968 he joined theRadar Systems group at Lin-coln Laboratory, and in 1972he was appointed assistantleader of the Digital SignalProcessing group. He wassubsequently appointed associ-ate leader and then leader of the newly formed SensorProcessing Technology group.In this capacity, he directedthe design and fabrication of several state-of-the-art radarsignal processors. He alsohelped identify the need forcustom integrated circuits withspecific application to radarsignal processing, and severalof these components were laterfabricated by the IntegratedCircuit group. He has pub-lished several articles on radarsignal processing, notably chapter 10 of Radar Technol-ogy , edited by E. Brookner,and chapter 5, with J.H.McClellan, of Applications of Digital Signal Processing Tech-nology , edited by A.V. Oppen-heim. He also published, withG. Heiligman, an article on an

adaptive beamforming proces-sor. For three years he servedas Radar Editor for the IEEE Transactions on Aerospace and Electronic Systems .

.

received S.B.E.E., S.M.E.E.,and Electrical Engineer degreesfrom MIT. He joined LincolnLaboratory in 1968 as a tech-nical staff member and wasappointed assistant leader of the Speech Systems Technol-ogy group in 1979. He wasappointed associate leader in1981, and became associatehead of the Computer Tech-nology division in 1984. In1995 he joined the Surveil-lance and Control division asan associate division head,

with responsibilities in theInformation System Technol-ogy area. In 1998 he wasappointed senior staff in theCommunications and Infor-mation Technology divisionoffice. Later that year he wasassigned as senior staff in theDirectors Office, with respon-sibilities in the areas of tech-nology transfer, intellectualproperty, continuing technicaleducation, and informationtechnology infrastructure. Hisresearch interests include high-performance computingarchitectures, digital-signal-processing algorithms andsystems, speech processingsystems, efficient digital sys-

tem implementations, novelapplications of VLSI and wafer-scale integration circuittechnology, and medicalelectronics. He is a member of Tau Beta Pi and Eta KappaNu, and a Senior Member of the IEEE. He received theIEEE Centennial Medal in1984 for service to the SignalProcessing Society, where heserved on the ConferenceBoard for nine years.

received a B.S. degree fromSeattle University in 1950 andan S.M. degree from MIT in1952, both in electrical engi-neering. After teaching atSeattle University for fouryears, he joined the MicrowaveComponents group at LincolnLaboratory. Of the microwavesystems he helped develop themost notable were used on theLaboratorys Haystack plan-etary radar in the fourth test of Einsteins general theory of relativity, and on the ALCOR radar at the Kwajalein Atollmissile test range, which iscapable of imaging reentry vehicles. In 1967 he becameassociate leader, and in 1968group leader, of a group thatdesigned, built, and testedcomplete prototype radarsystems. The first system was aradar used in Vietnam todetect people walking underdense foliage. Starting in 1972his group developed digitalsignal and data processorscapable of completely auto-matic detection, tracking, anddisplaying of moving targets inheavy clutter. This work led toa netted radar system demon-strated by the Army Artillery

at Fort Sill, Oklahoma, anairborne radar to detect slowly moving ground vehicles (theprogenitor of the Joint STARSradar), and the prototype of a

widely employed FAA AirportSurveillance Radar (ASR-9).For seven years he served asRadar Editor of the IEEE Transactions on Aerospace and Electronic Systems .


24/24


was associate head of the SolidState division from 1983 to1995. As leader of the AnalogDevice Technology group from1969 to 1983, he led andparticipated in the develop-ment of surface-acoustic-wave(SAW) and superconductive-device technology as well as X-ray lithography and precisionion-beam-etching equipment.The SAW activity producedunique, precisely matchedpulse compressors for the

ALCOR all-range channel andfor the multiple-antennasurveillance radar (MASR), as

well as enabling devices formilitary satellite (MILSAT)communication prototypes.

Additional devices were devel-oped and incorporated intoprototypes of high-perfor-mance spread-spectrum IFFsystems, a packet-radio system,and a missile-borne antijam-ming system. Before joiningLincoln Laboratory in 1964,he developed ferrite microwavedevices at Sperry Gyroscopeand General Electric Compa-nies, and was director of research at Microwave Chemi-cal Laboratory. He served inthe U.S. Navy from 1946 to

1948, received a B.S. degree inelectrical engineering fromColumbia University in 1953,and attended the GraduateSchool of Electrical Engineer-ing at Cornell University from1953 to 1955. He retired fromthe Laboratory in March1995.

.

is a senior staff member in theEmbedded Digital Systemsgroup. He received a B.E.E.degree and an M.E.E. degreein electrical engineering fromthe Polytechnic Institute of Brooklyn. His many accom-plishments include research onspeech bandwidth compres-sion, contributions to the fieldof digital signal processing,application of optical tech-niques to educational technol-ogy and communication, andinvestigation of space-basedradar systems. He has been atLincoln Laboratory since1961. From 1971 to 1982 he

was an assistant group leaderin the Spacecraft Processorsgroup, which built the LES-8and LES-9 communicationssatellites launched in March1976. He is a Fellow of theIEEE and past president of theIEEE Acoustics, Speech andSignal Processing (ASSP)Society. He has received the

ASSP Technical Achievement Award (1976), the ASSPSociety Award (1985), and theIEEE Jack S. Kilby Award(1996). He has also taughtcourses on advanced digitalsignal processing in China and

Mexico. His books includeDigital Processing of Signals (coauthored with Ben Gold),Number Theory in Digital Signal Processing (coauthored

with James McClellan), Ad-vanced Digital Signal Processing (coauthored with John G.Proakis, Chrysostomos L.Nikias, and Fuyun Ling) andDigital Signal Processing (coed-ited with Lawrence Rabiner).

.

is a senior staff member in theElectro-Optical Materials andDevices group, where hecarries out research and devel-opment of a wide range of optical devices and subsystems.He received S.B. and Ph.D.degrees in physics from MIT.His Ph.D. thesis work onultrasonic studies of low-temperature physics was fol-lowed by a five-year post-doctoral position in the MITMaterials Science Center. He

joined Lincoln Laboratory in1970 and worked on thedevelopment of SAW devicesfor signal processing in radarand communication systems.He led the development of thefirst reflective-array compres-sors and became associateleader of the Analog DeviceTechnology group. His transi-tion from basic research tosystems development includeda trip to the Kwajalein Atollradar-measurements field site,

where he debugged signalprocessing hardware andinstalled wideband SAW devices in the ALCOR radar.In 1980, he became the leaderof the Applied Physics group,

which carried out develop-

ment of optical devices. Hehas published three book chapters and many technicalpapers on a wide range of topics. Professional activitiesinclude considerable involve-ment with the Optical Society of America and the IEEE. Heis an IEEE Fellow, received aCareer Achievement Award,and was a National Lecturer.

• purdy, blankenship, muehe, stern, rader, and williamson

Documents