IEEE TRANSACTIONS ON COMPUTERS, VOL. c-24, NO. 11, NOVEMBER 1975 1074 [6] W. D. Strecker, "An analysis of the instruction execution rate in certain computer structures," Ph.D. dissertation, Dep. Elec. Eng., Carnegie-Mellon Univ., Pittsburgh, Pa. (Rep. AD-711408) June 1970. [7] M. Pirtle, '"Intercommunication of processors and memory," in 1967 Fall Joint Computer Conf., AFIPS Conf. Proc., vol. 31. Washington, D. C.: Thompson, pp. 621-633. [8] E. G. Coffman, Jr., "A simple probability model yielding performance bounds for modular memory systems," IEEE Trans. Comput. (Short Notes), vol. C-17, pp. 86-89, Jan. 1968. [9] G. J. Burnett and E. G. Coffman, "A study of interleaved memory systems," in 1970 Spring Joint Comput. Conf., AFIPS Conf. Proc., vol. 36. Montvale, N. J.: AFIPS Press, pp. 467474. [10] A. Parzen Stochuastic Processes. San Francisco, Calif.: HoldenDay, 1962. [11] W. R. Franta and P. A. Houle, "Comments on 'models of multi-processor multi-memory bank computer systems,'" in Proc. Winter Simulation Conf., Jan. 14-16, 1974, pp. 87-97. _ Kuchibhotla V. Sastry (S'71-M'72) was born in Vijayawada, India, on October 8, 1945. He received the B.E. degree from j Andhra University, India, in 1966, the M.E. from the Indian Institute of Science, degree Bangalore, India, in 1968, and the M.S. 7 and Ph.D. degrees from the University of Minnesota, Minneapolis, Minn., in 1970 and 1973, respectively, all in electrical engiFrom 1968 to 1973 he was first a Teaching Assistant and later a Reasearch Assistant in the Department of Electrical Engineering, University of Minnesota while working on his doctoral dissertation. He joined Sperry Univac, Roseville, Minn., in 1973 where he has been working on multiprocessor system organization and virtual memory management techniques. His current interests include computer architecture, operating system design, and modeling techniques. Dr. Sastry is a member of the IEEE Computer Society. Richard Y. Kain (S'56-M'58) was born in Chicago, Ill., on January 20, 1936. He received the B.S., M.S., and Sc.D. degrees in electrical engineering from the Massachusetts Institute of Technology, Cambridge, in 1957, 1959, and 1962, respectively. He has been an Associate Professor of Electrical Engineering at the University of Minnesota, Minneapolis, since 1966. Prior to that he was a Teaching Assistant, Instructor, and Assistant Professor at M.I.T. in the Department of Electrical Engineering. He has worked on time-shared computing, automata theory, computer architecture, and computer operating systems. He has published one book and numerous technical papers. Dr. Kain is a member of the American Association for the Advancement of Science, the Association for Computing Machinery, Sigma Xi, and Eta Kappa Nu. A/D Conversion Using Geometric Feedback AGC DENNIS R. MORGAN, Abstract-A digital signal processing technique is described which utilizes a built-in automatic gain control (AGC) function. A particularly attractive algorithm called "geometric feedback" is developed which has certain desirable properties. A simple analytic solution of the response is derived for the special case of linear geometric feedback. Index Terms-AID conversion, automatic gain control (AGC). I. INTRODUCTION TN many signal processing applications, the long-term dynamic range of the signal is too large to be accommodated using practical A/D converters. The use of automatic gain control (AGC) prior to A/D conversion is commonly used in this situation to reduce the long-term dynamic range. Short-term signal variations must still, of course, be within the range of the A/D converter. Manuscript received May 15, 1974; revised April 15, 1975. The author is with the Electronies Laboratory, General Electric Company, Syracuse, N. Y. 13201. MEMBER, IEEE Minicomputers are now commonly used in many realtime signal processing systems. In this case, it is desirable to use the vast amount of computational power that is available in order to minimize the external hardware required. The present correspondence is concerned with this aspect of design as it applies to AGC. A readily available component that is particularly useful for digital AGC is the multiplying D/A converter (MDAC). A block diagram of a signal processing system using this component in an AGC loop is shown in Fig. 1. Here a digitally controlled attenuator is formed by using the MDAC in the feedback loop of an op-amp. The attenuation sequence that controls the MDAC is computed from the input samples using some algorithm. A particularly attractive algorithm called "geometric feedback" will be developed in this correspondence which has certain desirable properties. In particular, the feedback function can be tailored to give the desired largesignal transient response while the small-signal bandwidth remains independent of input level. MORGAN: A/D CONVERSION 1075 USING FEEDBACK AGC GAIN CONTROL OUTPUT MINICOMPUTER ANALOG SIGNAL DATA INPUT DIGITALLY-CONTROLLED ATTENUATOR Fig. 1. Block diagram of signal processing system with digital AGC. II. ANALYSIS Fig. 2 shows an envelope equivalent block diagram of a discrete-time AGC amplifier where x(n) is the input and y(n) the output. For convenience, the inputs and outputs are assumed to have been scaled to a reference of unity with no other gain factors appearing in the loop. An error sequence e(n) is derived by subtracting the reference from the output and is filtered by an accumulator. The accumulator output b (n 1) controls a nonlinear attenuator which is represented by the function f(b) and a divider. An alternative configuration of Fig. 2 could, of course, be arranged by replacing the divider with a multiplier and replacing f with l/f. The function f is assumed to be a continuous, strictly monotonically increasing function.' The system of equations describing the model is given by - y(n) = x(n)/a(n a (n) = f[b (n) ] (ib) b(n) = b(n (lc) e(n) = y(n)-1 - - 1) (la) 1) + e(n) and (1d) The time index n for these variables implies discrete samples taken every T seconds. It can be shown by linearization that for a nominal trajectory x== = x and yn 1, the normalized smallsignal transfer function of Fig. 2 is given by [1] small-signal time constant is defined as for O < < 1. T = -T/ln (5) As can be seen from the transfer function, the AGC amplifier is stable in a small-signal sense provided that a < 1, i.e., the incremental loop gain satisfies the condition O < K < 2. As in the continuous case, it is noted that for exponential feedback f(b) = exp (yb), the loop gain K = is constant and hence the small-signal response is independent of input level [1]-[3]. This is a desirable characteristic for many applications. Besides maintaining a levelindependent frequency response, exponential feedback also affords an extra measure of protection against conditional instability since the loop gain is constant. For exponential feedback, (lb) and (lc) can be combined as a(n) = a(n - 1) exp [Eye(n) ]. (6) This suggests the alternative configuration shown in Fig. 3 which is equivalent to Fig. 2 for g(e) = f(e) = exp (,ye). Since a constant error would give rise to a geometric attenuation sequence, this configuration is deemed "geometric feedback." The function g is required to be strictly positive, monotonically increasing, and in addition g(O) = 1. For steady-state conditions, it is easy to see that the incremental loop gain is given by a, a y K g'[g-(1)] (7) z1 H(z) = Z aE *(2) and the small-signal relationships (2), (3), and (5) all hold with (4) replaced by (7). It is significant that (7) is independent of signal level, a desirable property as prewhere viously mentioned. If g(e) = f(e) = exp (eye), then (4) a1-K (3) and (7) are of course equivalent. The significance of the geometric feedback configuraand tion is that the function g can be tailored to give the K = f'[fE () ]/x > 0. (4) desired large-signal transient response while the smallThe "loop gain" K and hence the transfer function are signal response remains independent of input level. An alternative configuration of Fig. 3 could be arranged seen to vary according to the average input level x. A by replacing the divider with a multiplier and replacing 'Alternatively, a monotonically decreasing function can be used g with i/g. This configuration may be more desirable for if the polarity of the error signal is reversed, i.e., e(t) 1 y(t). implementation in some cases. - = - = 1076 IEEE TRANSACTIONS ON COMPUTERS, NOVEMBER 1975 Fig. 2. Envelope equivalent circuit of discrete AGC amplifier with accumulator filter. (u)X~~~ a E (T U) Fig. 3. Envelope equivalent circuit of discrete geometric feedback AGC amplifier. y(n) 1.0 0.1 L 0 1 2 3 4 5 6 7 nT/T Fig. 4. Linear geometric feedback AGC amplifier response to step input. The general system equations are y(n) = x(n)/a(n - 1) (8a) linear difference equation a(n + 1) - a(n) + E[x(n + 1) - a(n)] (10) which is the discrete version of the continuous case with exponential feedback [1]. The z transform of the solution of (10) is (8b) a(n) = a(n - l)gre(n)] (8c) e(n) = y(n),- 1 and the difference equation for the attenuation is z (11) A (z) = X(z) = z- a (9) a(n + 1) a(n)g[x(n + 1)/a(n) 1]. If g is a linear function of the form g(e) = 1 + -ye (the where, as before, a = 1 - and the same conditions apply first-order expansion of exp (,ye)), then (9) becomes the regarding the effect of y on stability. MORGAN: A/D 1077 CONVERSION USING FEEDBACK AGC 0 0 0~~~~ _ 0 o.1 o N - : o 0 0.001. nT/T response to step input for linear geometric Figg. 5. Normalized error feedback AGC amplifier. For a step input x(n) = In practice, it is necessary to limit 1 + ye so that g > gqin > 0, for all e. A plot of normalized error is shown in Fig. 5 and demonstrates the asymptotic exponential behavior predicted from the small-signal theory. For other choices-of g, the response will generally involve a nonlinear difference equation. Methods for analyzing this situation using the discrete Volterra series are described in [1]. x(O), x(O) /yo, n <0 the solution of (11) in terms of y is I( - (= - 1/yO) an]- n >0 (12) III. CONCLUSIONS A new digital AGC algorithm has been developed which where yo = y(0) specifies the initial output level and is is called "geometric feedback." This algorithm has the equal to the ratio of the input step transition. Equation desirable property that the small-signal response is inde(12) is plotted in Fig. 4 for various input levels. As can pendent of input level. This property holds independent be seen convergence is faster for yo < 1 but slower for of the particular nonlinearity used. Therefore, the nonyo > 1 as would be expected since exp (ye) > 1 + ye. linearity can be tailored to give the desired large-signal 1,1 n <0 1078 IEEE TRANSACTIONS ON COMPUTERS, VOL. transient response while preserving level invariant smallsignal response. In the simplest form, a linear feedback function results in a linear difference equation which can be solved in terms of an arbitrary input. Level-invariant small-signal response was demonstrated by illustrating the asymptotic convergence of the error to an exponential decay. Other types of nonlinearities in the geometric feedback algorithm are a subject for future investigations. NO. 11, NOVEMBER 1975 [3] A. G. Morris, "A constant volume amplifier covering a wide dynamic range," Electron: Eng., vol. 37, pp. 502-507, Aug. 1965. Dennis R. Morgan (S'63-S'68-M'69) was born in Cincinnati, Ohio, on February 19, 1942. He received the B.S. degree in 1965 from the University of Cincinnati, Cincinnati, Ohio, and the M.S. and Ph.D. degrees from Syracuse University, Syracuse, N.Y., in 1968 and 1970, respectively, al in electrical REFERENCES [1] D. R. Morgan, "On discrete-time AGC amplifiers," IEEE Trans. Circuits Systems, vol. CAS-22, pp. 135-146, Feb. 1975. [2] W. K. Victor and M. H. Brockman, "The application of linear servo theory to the design of AGC loops," Proc. IRE, vol. 48, pp. 234-238, Feb. 1960. c-24, engineering. He is a Senior Engineer in the General Electrical Company, Electronics Laboratory, Syracuse, N.Y., where he is involved with the analysis and design of signal processing systems. System Fault Diagnosis: Closure and Diagnosability with Repair JEFFREY D. RUSSELL, MEMBER, IEEE, AND Abstract-Determination of the detectability and diagnosability of a digital system containing at most t faulty system components is considered. The model employed is to an extent independent of the means used to implement diagnostic procedures, i.e., whether the tests are accomplished via hardware, software, or combinations thereof. A parameter, called the closure index, is defined which characterizes the capability for executing valid tests in the presence of faults. The closure index can be thought of as the size of the smallest potentially undetectable multiple-fault in the system as modeled. On the basis of this parameter, results are presented which permit the determination of t-fault detectability and t-fault diagnosabiity with repair for the system. Examples are presented to illustrate the application of the model for systems close to those encountered in actual practice. Index Terms-Closure index, diagnostic modeling, diagnosable digital systems, fault detection and diagnosis, fault-tolerant computing. Manuscript received April 4, 1974; revised April 29, 1975. This work was supported in part by the National Science Foundation under Grants GK-3844 and GK-35037. J. D. Russell was with the Department of Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, Wis. 53706. He is now with the Collins Avionics Division, Rockwell International, Cedar Rapids, Iowa 52406. C. R. Kime is with the Department of Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, Wis. 53706. A portion of this research was performed while he was with the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, Calif. CHARLES R. KIME, MEMBER, IEEE INTRODUCTION SEVERAL studies [1]-[9] have been concerned with system-level fault diagnosis and have had as an objective the determination of conditions for diagnosability. Interest in this topic is motivated by the need for highly available digital systems that can continue operation, perhaps at reduced capacity, when multiple hardware failures occur. The approach to such systems via reconfiguration or standby sparing requires that the presence of malfunctioning elements be detected and their location determined to within a system module, i.e., the system must be diagnosable. Moreover, initial diagnosis to the level of large replaceable modules can reduce system downtime in those cases in which manual repair is performed. Most of the previous models for system-level diagnostic analysis have considered the system to be partitioned into a number of disjoint subsystems under the assumption that each subsystem can be completely tested by some combination of the other subsystems. Each test so defined involves the controlled application of stimuli to the subsystem under test and the analysis of the ensuing responses, resulting in the evaluation of the tested sub-