Recent advances in analog signal processing.
The recent advances in the design and application of analog signal processors for implementing functions such as convolution, correlation, matched filtering and Fourier transformation are described. The performance of programmable devices which can handle the changing waveform types used in sophisticated communication and radar systems are featured.
Most of the signal-processing operations considered here can be performed by devices having the structure shown in Figure 1. When the lower outputs are all combined at a single node, the structure represents the finite-impulse-response (FIR), or transversal, filter whose impulse response is controlled by the stored weight values [W.sub.n].
The alternative approach to an FIR matched filter receiver is the active correlator used in spread-spectrum communication receivers, which is identical to the single-range-cell time-integrating correlator used in coded radar systems. This is realized by reducing Figure 1 to only the first delay and multiply stage, with the output terminated in an integrator.
To implement correlation, the reference weight [w.sub.1] must not be altered in synchronism with the input signal, which restricts the operation of this device to applications where the signal time-of-arrival is known in advance. This deficiency can be alleviated by applying the same reference weight to all the correlators in Figure 1, with each one terminated in a separate integrator. This processor realization is referred to as the time-integrating correlator. Analog realizations of Fourier-transform processors often make use of convolution operations which are performed in structures similar to that shown in Figure 1.
The basic components which are used to realize the delay, multiply, etc., functions in these different analog devices are summarized in Table 1. [Tabular Data Omitted]
SIGNAL PROCESSING DEVICES
A. Charge-Coupled Devices
Sampled-data analog programmable transversal filters (PTFs) have been available for the last decade as CCD and bucket-brigade devices, with CCDs now being the preferred type.
For full analog-analog correlation PTF capability, two distinct CCD approaches have been developed. In the first approach direct mixing of the stored analog signal samples with the analog filter weight can be accomplished in a single metal-oxide-semiconductor (MOS) transistor multiplier at each tapping point. Alternative designs use a four-FET (field-effect transistor) bridge circuit for increased accuracy.
When combined with the CCD tapped delay line, these compact realizations of the transversal filter cell allow 256-point PTFs to be designed in a single integrated device. This approach thus offers compact, low-power PTFs, but the difficulties in performing analog-analog operation with sufficient accuracy limit the device to clock rates of 1 to 5 MHz and hence to usable signal bandwidths of several hundred kilohertz for programmable frequency filtering and matched filtering of linear-frequency-modulated (LFM) chirp waveforms.
B. Surface Acoustic Wave Devices
In many spread-spectrum systems, the bandwidth of the signal that must be processed is closer to 100 MHz than 1 to 25 MHz, precluding the use of CCD sampled-data processors. However, these wide bandwidths can readily be accommodated in continuous-time analog SAW processors. As with CCDs, SAW convolvers, correlators and matched filters, take many distinct designs.
All convolvers are attractive, as they use a simple construction without any structurally imposed coding and the programmability is achieved by varying the modulation on the time-reversed reference waveform. Thus they can function as either an analog-binary correlator or full PTF which can be programmed in the time it takes a new reference waveform to propagate into the device (1 to 50 [mu] sec). Further, the lack of structurally imposed reference-waveform coding makes the device adjustable to minimize Doppler effects.
C. Superconductive Devices
Exceeding the bandwidths of CCD and SAW matched filters, superconductive electromagnetic components have demonstrated several advanced filter functions at > 1 GHz bandwidth. The first device developed, the superconductive LFM pulse-compression filter, is based on a coupled pair of superconductive spiral striplines, surrounded by two layers of low-loss dielectric which are coated with superconductive niobium ground planes.
These devices are again weighted tapped-delay-line structures as shown previously in Figure 1, but for a dispersive filter design the individual delay terms each have a different value. Energy is tapped or coupled from the input delay line to the output line through a cascade of backward-wave couplers which individually provide peak coupling at frequencies for which the coupler length equals an odd number of quarter wavelengths. The strength of individual couplers is controlled by the spacing between the striplines within the coupler and can be tailored for the usual weighting functions.
Current work is aimed at improving dimensional tolerances and packaging to achieve a value of < -40 dB, which is comparable with SAW device performance, and at achieving the thin (10 [mu] m) low-loss dielectrics required to reach delays of 1 [mu] sec.
Superconductive processing can also be extended to convolvers by sending signals into opposite ends of a superconductive delay line and sampling the counterpropagating signals with weakly coupled proximity taps which are connected to superconductor-insulator-superconductor (SIS) tunnel-junction mixers. These SIS mixers are arranged in balanced series arrays to enhance the desired cross products while suppressing the self products and higher-order terms. With provision of a time-reversed reference, this convolver can perform as a multi-gigahertz-bandwidth device.
D. Optical and Acousto-optic Processors
Most optical processors are more specifically acousto-optic processors, relying on Bragg cells to convert the electronic input data to optical form. The acousto-optic Bragg cell is thus a particularly significant component, both on its own as a Fourier-transform processor and as the vital modulator in acousto-optic correlators and convolvers.
In the Fourier analyzer a propagating acoustic wave, either in the bulk or on the surface, interacts with an optical beam to give a set of diffracted optical outputs whose angular deflection is proportional to the frequency of the acoustic wave. For a wideband acoustic-wave input, the deflected signal magnitudes are controlled by the amplitude of the spectral components. Analysis is performed by focusing the difracted optical output onto a photodiode array which may incorporate CCDs for serial readout capability. Frequency resolution is approximately equal to the reciprocal of the acoustic transit time through the optical beam. Engineered versions of both the bulk-wave and surface-wave implementation of this radiometric device now exist.
The current performance capabilities of wideband acousto-optic spectrum analyzers for radar electronic support measures (ESM) are 1-2 GHz bandwidth, 2-20 MHz resolution and a time-bandwidth product or number of resolution cells approaching 1,000. In fully engineered systems with associated photodetector arrays, the number of resolution cells is typically 25-200, with the lower values applying to fiber coupling of individual detectors. For communications ESM applications, similar time-bandwidth products can be achieved with 20-50 kHz resolution, but this requires acousto-optic substrate material with much slower acoustic velocity.
Acousto-optic spectrum analyzers in principle offer exceptionally fast response, as they operate in the time it takes the acoustic wave to transit through the processor. However, the efficiency of the photodetector is a problem, as it has to integrate for tens of microseconds or even milliseconds to give satisfactory output voltage.
E. Comparison of Analog and Digital Processors
The performances of selected devices described above are summarized in Table 2, including analog-binary correlators for the detection of binary phase-coded waveforms and full analog-analog correlators which act as PTFs to achieve arbitrary filter functions. This table classifies the current parameters in terms of bandwidth, delay and time-bandwidth product and indicates the performance limitations separately in terms of noise-limited dynamic range and spurious-response or accuracy limits. [Tabular Data Omitted]
As variable-clock components, a given CCD design can in principle be operated over greater than 100:1 range of absolute center frequency and bandwidth, giving the very wide figures included in Table 2. The maximum delay of the processed signal is restricted by dark-current effects at low clock rates. As SAW components are usually optimized for large bandwidths, the minimum bandwidth is specified here as 5 MHz. The range of figures provided in the time-delay column has been adjusted to correspond to the minimum and maximum bandwidths specified in the previous column. Thus the achievable processor delay in wideband SAW devices is more restricted than narrowband designs. Only typical values are included for the newer superconductive devices.
The equivalent speed or computational throughput rate of the analog processors can be derived by calculating the effective number of arithmetic operations (real multiplications and additions) for an equivalent digital processor.
Digital processors normally implement convolution by frequency-domain multiplication, and either a radix-2 FFT or simpler DFT-based implementation is selected for comparison with analog processors as appropriate. As there are 10 basic arithmetic operations in an FFT butterfly computation and (N/2) [log.sub.2] (N) butterflies per N-point FFT, the computational rate in operations per second (OPS) required to perform real-time Fourier transformation of continuous signals is B[5a [log.sub.2] (aTB)] where B is the signal bandwidth, T is the data block duration and a is the ratio of the Nyquist-limited bandwidth (half the sampling rate) to the actual signal bandwidth B.
Figure 2 provides, for the key devices described here, a comparison of analog devices against various digital signal-processor chips, board processors and complete systems, based on the above equation. It also includes an indication of the computational rates which are typically required in a number of radar, sonar and image-processing systems. Note that even low-bandwidth sonar systems require high throughput rates because of the large number of parallel processing channels and the need to simultaneously process with different detection and localization algorithms.
Two factors must be considered when examining Figure 2. First, the analog techniques offer wide bandwidth. However, their accuracy due to spurious responses (Table 2) is only comparable with 6-8 bits of quantization at the output of an equivalent digital processor. Higher accuracy input quantization and precision of internal arithmetic, e.g., 10-14 bits in fixed point processors, is normally required to support this level of performance at the output terminal.
Many of the digital examples provided in Figure 2 are floating-point processors with typically 32-bit or higher precision arithmetic whose accuracy and dynamic range exceed the analog-device capabilities but at a much lower processor bandwidth. Thus there is a distinct bandwidth/accuracy trade-off between analog and digital processors.
With all signal processors, there is a performance benefit in an application-specific processor compared with a general-purpose design. The programmable digital convolver provides an example of a sophisticated processor, developed for a particular function. It is a 49-circuit-card design based on 22-bit floating-point arithmetic, using commercially available VLSI circuit elements with a total dissipation of 5.7 kW. It thus provides a computational rate which considerably exceeds general-purpose commercial digital array-processor capabilities.
Further, it comes close to the capabilities of the mid-'70s design of a sophisticated digital radar processor which used less accurate custom ECL circuits, dissipated much more power and occupied a set of 8-ft-high cabinets rather than 10 cubic feet. However, the sophisticated digital convolver which is optimized for one function still lags behind the computational performance of the SAW convolver.
It is obvious that digital techniques will continue to dominate in applications where accuracy and flexibility are more important than the bandwidth, peak processing load, size, weight and power consumption of the processor. The system designer needs to make a careful decision on whether to select an analog or digital implementation.
It is confidently predicted that application-specific analog processors are significant in modest-bandwidth and modest-accuracy applications where size and power are highly valued. CCDs and lower-bandwidth SAW devices, which provide low unit cost in high-volume applications, clearly fall into this category. The modest-bandwidth analog components reported here can also be used for front-end processing, followed by low-cost, reduced accuracy A/D conversion, as demonstrated in the spread-spectrum receiver. In some processors digital integration will follow, yielding a reduced data rate output with higher accuracy than is obtainable directly from analog components themselves.
Wide-bandwidth signal processing systems are possible but not attractive digitally, as physically immense processors would be required. Because they directly exploit physical phenomena and are structurally much simpler than digital processors, analog processors will always offer an order of magnitude more bandwidth than digital processors for computationally intensive tasks.
However, the problem of acquisition and interfacing of the output data to subsequent digital processors does make it difficult to quickly apply the extremely high bandwidth superconductive technology, except for the time-integrating devices. Other significant application areas for analog processors occur, for example in image processing, where computationally intensive preprocessing operations can be integrated within the analog sensor.
In summary, CCDs merge conventional fabrication technology with parallel architectures to achieve large processing power with small physical size and low electrical power. Wider-bandwidth devices harness physical phenomena which are inherently analog (such as SAW) or, as in the case of superconductivity, which have not yet reached the level of fabrication refinement necessary to produce digital processors with high yield. It is highly unlikely that even future VHSIC-based processors will achieve the high signal bandwidths offered by wideband SAW and superconductive analog components. However, with more and more engineers specializing in computing and digital techniques, the development of new and the further refinement and application of existing analog devices is becoming an increasingly challenging task for the limited numbers of circuit designers devoted to analog design.
PHOTO : FIGURE 1 BASIC ARCHITECTURE FOR SIGNAL PROCESSOR FUNCTIONS SUCH AS FIR FILTER AND CORRELATOR
PHOTO : FIGURE 2 COMPUTATIONAL SPEED OF ANALOG AND DIGITAL SIGNAL PROCESSORS IN COMPARISON WITH TYPICAL SYSTEM REQUIREMENTS. COMPUTATIONAL RATE IS CLASSIFIED IN REAL ARITHMETIC OPERATIONS (E.G., ADD, MULTIPLY, ETC.) PER SECOND OF FFT-BASED IMPLEMENTATION.
Peter M. Grant is currently a visiting assistant professor at the Ginzton Laboratory, Stanford University. Richard S. Withers is the associate group leader in the Analog Device Technology Group at MIT Lincoln Laboratory.
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||EW Design Engineers' Handbook & Manufacturers Directory|
|Author:||Grant, Peter M.; Withers, Richard S.|
|Publication:||Journal of Electronic Defense|
|Date:||Jan 1, 1992|
|Previous Article:||Space power systems requirements and issues: the next decade.|
|Next Article:||Engineering rules and design charts.|