Sunteți pe pagina 1din 6

7

Modeling Interconnected Systems:


A Functional Perspective
Stanley R. Liberty 7.1 Introduction ....................................................................................... 1079
Academic Affairs, Bradley University, 7.2 The Component Connection Model ........................................................ 1079
Peoria, Illinois, USA 7.2.1 Detailed Insight to C o m p o n e n t Connection Philosophy
7.3 System Identification ............................................................................ 1082
7.4 Simulation .......................................................................................... 1082
7.5 Fault Analysis ...................................................................................... 1082
7.6 Concluding Remarks ............................................................................ 1083
References .......................................................................................... 1083

7.1 Introduction pressed for simplicity of presentation in this chapter.) This


may be abstractly notated by:
This chapter, adapted from Liberty and Saeks (1973, 1974),
contains a particular viewpoint on the mathematical modeling y = Su, (7.1)
of systems that focuses on an explicit algebraic description of
system connection information. The resulting system repre- where S is the symbol for the input/output relation (the
sentation is called the component connection model. The system). On the other hand, one may think of the physical
component connection approach to system modeling provides system symbolized by S as an interconnection of components.
useful ways to tackle a variety of system problems and has Engineering experience explains that a particular S is deter-
significant conceptual value. The general applicability of this mined by two factors: the types of components in the system
viewpoint to some of these system problems is presented here and the way in which the components are interconnected. This
at a conceptual level. Readers interested in specific examples latter observation on the physical structure of a system gives
illustrating the component connection model should first refer rise to the component connection viewpoint.
to Decarlo and Saeks (1981). Additional references are pro- Again, thinking of a physical system, the system's connection
vided on specific applications of the component connection structure can be held fixed while the types of system compo-
philosophy to system problems. Readers with a clear under- nents or their values change. Experience explains that for each
standing of the general concept of function should have no distinct set of components or component values, a unique
difficulty comprehending the material in this section. input/output relation is determined. From a mathematical
perspective, a given connection description determines a map-
ping of the internal component parameters to the input/
7.2 The Component Connection Model output relation.
Now, one can generalize this slightly by thinking of the
A system, in simplistic mathematical terms, is a mapping from physical system as an interconnection of subsystems, each
a set of inputs (signals) to a set of outputs (signals) (i.e., an with its own input/output relation. Note that a subsystem
input/output relation with a unique output corresponding to may be either a discrete component, an interconnection of
each input). (In using this external system description, the components, or an interconnection of "smaller" subsystems.
context of an internal system state has been deliberately sup- This hierarchical structure is germane to the component con-

Copyright © 2005 by AcademicPress. 1079


All rights of reproduction in any form reserved.
1080 Stanley R. Liberty

nection philosophy and is precisely the fundamental concept tions, then one seeks a left inverse off. Here, the concern may
of large-scale system theory. be with uniqueness, and fis viewed as one-to-one and into the
Thinking in this way, one notes that a given connection set of input/output relations.
description determines a system-valued function of a system- Another classical system problem encountered in design
valued variable. Because this function is entirely determined by analysis is that of determining the sensitivity of the input/
the connections, it is called the connection function: output relation S to variations in certain component values.
In this case, differentiation of the connection function fwith
S = f(Z), (7.2) respect to the component parameters of interest is the major
constituent of the sensitivity analysis. Specific applications of
where Z symbolizes the internal subsystem descriptions or the component connection model to sensitivity analysis are
component parameters and where fsymbolizes the connection contained in Ransom ( 1973, 1972), Ransom and Saeks ( 1975),
function. In the special case where the subsystems represented and DeCarlo (1983).
by Z are linear and time-invariant, Z may be interpreted as a The conceptual description just described should provide
matrix of component transfer functions, gains, or immitances the reader with a "feel" for the component connection phi-
(the precise physical interpretation is unimportant), which is losophy. This chapter now provides more detailed insight.
also the case for S. One should not, however, restrict the
thinking process to linear time-invariant systems. In addition,
7.2.1 Detailed Insight to Component
it is important to point out that, even when all of the subsys-
Connection Philosophy
tems are linear, the connection function fis nonlinear.
What may be surprising is that for most commonly encoun- Although an abstract functional interpretation of the con-
tered systems, the nonlinear function f can be completely nections in a system is conceptually valid, it is of no practical
described by four matrices that permit the connection function value unless one has a specific and computationally viable
to be manipulated quite readily. Indeed, feasible analytical and representation of the function. Fortunately, there is such
computational techniques based on this point of view have a representation, and the conceptual problem of inverting
been developed and applied to problems in system analysis or differentiating the connection function may actually be
(Singh and Trauboth, 1973; Prasad and Reiss, 1970; Trauboth carried out.
and Prasad, 1970), sensitivity studies (Ransom, 1973, 1972; Classically, in circuit and system theory, connections are
Ransom and Saeks, 1975; DeCarlo, 1983), synthesis (Ransom, represented by some type of graph (e.g., linear graph, bond
1973), stability analysis (Lu and Liu, 1972), optimization graph, signal flow graph) or block diagram. This graphical
(Richardson, 1969; Richardson et al., 1969), well-posedness depiction of connection information can readily be converted
(Ransom, 1972; Singh, 1972), and fault analysis (Ransom, into a set of linear algebraic constraints on the circuit or system
1973, 1972; Saeks, 1970; Saeks et aL, 1972; Ransom and variables for subsequent analysis. As such, it is natural to adopt
Saeks, 1973a, 1973b; DeCarlo and Rapisarda, 1988; Reisig an algebraic model of the connections from the start.
and DeCarlo, 1987; Rapisarda and DeCarlo, 1983; Garzia, The precise form of the component connection model with
1971). The reason for this core of linearity in the fdescription its algebraic description of a system's connection information
is that connection information is generally contained in "con- may be seen by examining Figure 7.1. Figure 7.1(A) shows a
servation" laws that, upon the proper choice of system vari- box depiction of a system with inputs u and outputs y. The
ables, yield linear "connection" equations. inner box labeled Z represents the system components, and the
To obtain an intuitive feel for what types of manipulations outer donut shaped area represents the system connections
one performs on f to solve the systems problems previously (e.g., scalers, adders, Kirchhoff law constraints, etc.). Now, if
mentioned, consider, for example, a classical passive network Z represents a matrix of component input/output relations,
synthesis problem. In this problem, one is given the input/ the following may be abstractly written:
output relation S, and one knows what type of network struc-
ture is desired. For example, a ladder synthesis problem state- b = Za, (7.3)
ment consists of a transfer function specification and the
"connection information" of ladder structure. The S and fare where b denotes the vector of component output variables and
specified, and it is desired to determine Z (the components). a the vector of component input variables. The component
Clearly, to find Z, one must invert f. In this case, one seeks a connection model can be obtained by re&awing the system of
right inverse of f because the uniqueness of Z is not a concern Figure 7.1(A) as in Figure 7.1(B), where the components and
and f is viewed as "onto" the class of transfer functions of connections have been separated. Thus, the overall system is
interest. composed of two interconnected boxes. One box contains the
If, on the other hand, one is attempting to identify the value components as described by equation 7.3, and the second
of a component in an interconnected system of components donut-shaped box contains the connections. The donut-
given external measurements and knowledge of the connec- shaped box has inputs u and b and outputs a and y. Finally,
7 Modeling Interconnected Systems: A Functional Perspective 1081

y = u

(A) System with Inputs / Outputs (B) Separated Components and Connections

FIGURE 7.1 A System as an Interconnection of Components

since the connections are characterized entirely by linear alge- fault isolation problem. Existence conditions for the L matrices
braic constraints, the donut-shaped box may be mathemati- were first studied by Prasad (Prasad and Reiss, 1970) and later
cally modeled (i.e., the connections) by the matrix equation: by Singh (1972) who have shown that the existence of L
matrices is a reasonable assumption. This is essentially the
[;] = r Lll L121 same type of assumption that one makes when assuming that
LL21 L22J [ : ] (7.4) the state equations of a system exist. As seen above, if the
components in the system are integrators, then the L matrices
The component connection model is thus a pair of equations, are precisely the state matrices.
7.3 and 7.4, with the matrix Z representing the components, The advantage of the component connection model over the
and the set of four L matrices representing the connections. classical graphical and diagrammatic connection models is
An interesting observation can be made here. In the very that, being inherently algebraic, the component connection
special case where each component is an integrator, the model is readily manipulated and amenable to numerical
following is true: techniques. Moreover, this model unifies the various graphical
and diagrammatic models. In fact, electronic circuits, which
b=a. (7.5) are commonly described by hybrid block diagram-linear graph
models, are handled as readily as passive circuits or analog
Substituting equation 7.5 into 7.4 yields: computer diagrams [see DeCarlo and Saeks (1981) for some
simple examples].
Using the component connection model, the desired repre-
= L l l b q - L12u.
(7.6) sentation for the connection function is obtained. This is
y = L21bq- L22u. illustrated by a linear example since the concept can be more
intuitively understood in the linear case. It is important to
The reader will recognize these equations as the familiar "state point out again, however, that more generality is actually
model" of linear dynamical system theory. Thus, intuitively, the present. Viewing Z as a matrix of transfer functions, the overall
component connection model may be viewed as a generaliza- system transfer function is desired. Simultaneous solution of
tion of the linear state model. In fact, the L matrices describe equations 7.1, 7.3, and 7.4 yields:
how to interconnect subsystems to form a large-scale system S,
just as integrators used to be patched together on an analog S = L22 q- L21(1 - ZLH) 1ZL12= f ( Z ) , (7.7)
computer to simulate a system. If one views the L matrices as a
generalization of the state model, where the integrators have and we have a representation of the connection function in
been replaced by general components, a beneficial cross fertil- terms of the four L matrices even though the connection
ization between component connection theory and classical function is nonlinear in Z. This matrix representation facili-
system theory results. It is possible and indeed useful to talk tates manipulation of the connection function as shown later
about controllable or observable connections (Singh, 1972). (Ransom, 1973, 1972; Saeks, 1970; Ransom and Saeks, 1973b,
The component connection model was first used intuitively 1973a). The first observation that the connection function
by Prasad and Trauboth (Singh and Trauboth, 1973; Prasad could be so represented was made by Saeks (1970), while the
and Reiss, 1970; Trauboth and Prasad, 1970) and formalized by first exploitation of the concept is attributable to Ransom
Saeks (Saeks, 1970; Saeks et al., 1972) for application in the (1973, 1972) and Saeks (1973b, 1973a).
1082 Stanley R. Liberty

7.3 System Identification tion viewpoint, the starting point is a specified collection of
decoupled component differential equations and a set of
In a system identification problem, one is normally asked to coupled algebraic connection equations. Thus, it is possible
identify the mathematical input/output relation S from input/ to take advantage of the special form of the equations charac-
output data. In theory, this is straightforward if the system is terizing a large-scale dynamical system and to formulate
truly linear (Chang, 1973), but in practice this may be difficult analysis procedures that are more efficient than those which
due to noisy data and finite precision arithmetic. The non- might result from a purely numerical study of the coupled
linear case is difficult in general because series approximations differential equation characterizing the entire system.
must be made that may not converge rapidly enough for large Such a procedure, namely a relaxation scheme, wherein one
signal modeling (Bello et al., 1973). In both linear and non- integrates each of the component differential equations by
linear identification techniques, random inputs are commonly separately iterating through the connection equations to
used as probes (Bello et al., 1973; Cooper and McGillem, 1971; obtain the solution of the overall system, was implemented
Lee and Schetzen, 1961). The identification of S is then carried over 30 years ago at the Marshall Space Flight Center (Prasad
out by mathematical operations on the system input and and Reiss, 1970; Trauboth and Prasad, 1970; Saeks, 1969). The
output autocorrelations and the cross-autocorrelations be- feasibility of the scheme was verified, and significant computer
tween input and output. memory savings resulted. The key to this approach of the
This section does not contain a discussion of such tech- numerical analysis of a large-scale system is that numerical
niques, but some observations are presented. It should be routines are applied to the decoupled composite component
noted that if only input/output information is assumed differential equations rather than an overall equation for the
known (as is the case in standard identification techniques), entire system. Thus, the complexity of the numerical compu-
then system identification does not supply any information on tations is determined by the complexity of the largest compo-
the system structure or the types or values of components. In nent in the system. In contrast, if the numerical procedures are
such cases, the component connection model cannot be used. applied directly to a set of differential equations characterizing
In many applications, however, connection information and the entire system, the complexity of the numerical computa-
component type information are available. In such cases, the tions is determined by the complexity of the entire system (the
component connection model may be applicable. Indeed, for sum of the complexities of all the components).
identification schemes using random inputs, it is easily shown Since 1970, computer simulation software has evolved sub-
that the component correlation functions and the overall stantially, and packages that deal with large-scale systems
system correlation functions are related by: either explicitly or implicitly exploit the component con-
nection philosophy. One of the more recent software de-
Ryy = L21abbLT1 + RyuLT22 + L22Ruy -- L22RuuL2T2. (7.8) velopments for modeling and simulation of large-scale
heterogeneous systems is an object-oriented language called
Raa = LllRbbL1T1 + LllRbuL~ + L12RubLTll + L12RuuL1T2. (7.9) Modelica; see Mattson et al. (1998) for additional information.

L21Rbu = [Ryu - L22Ruu]. (7.10)

Rab = LllRbb + L12Rub. (7.11)


7.5 Fault Analysis
The Rjk is the crosscorrelation of signal j and signal k. If the left Fault analysis is similar to system identification in that the
inverse of L21 exists, then the subsystem correlation functions system is reidentified to determine if the component input/
can be determined and used to identify the subsystems via output relations have changed. It should be clear that connec-
standard identification techniques. However, in general, this tion information is essential to detection of a faulty compo-
left inverse does not exist and either the type information or nent. As before, note that fault analysis is not restricted to
the psuedo-inverse of L21 must be used to yield subsystem linear systems. To illustrate some of the technique, the discus-
models. sion will focus on a linear time-invariant situation.
To attack the fault analysis problem, assume that there is a
given set of external system parameters measured at a finite set
7.4 Simulation of frequencies S(~oi), i = 1, 2 . . . . . n and that there are the
connection matrices, Lll, L12, L2~, and L22. Compute or ap-
There are numerical challenges in computer simulation of proximate Z(o~). For simplicity of presentation, assume that
large-scale systems. These difficulties are due to the simulta- L22 = 0. There is no loss of generality with this assumption
neous solution of large numbers of differential equations. The since one can always replace S(~oi) by S(coi) = S(o~i) - L22 and
efficiency in solving a single differential equation is entirely then work with the system whose measured external param-
determined by the numerical techniques employed. However, eters are given by S and whose connection matrices are
in analyzing a large-scale system from the component connec- Lll = Lll, L12 = L12, L21 = L21, and L22 = 0.
7 Modeling Interconnected Systems: A Functional Perspective 1083

In the most elementary form of the fault analysis problem, and utility of the philosophy. Even though the component
assume that S(~o) is given at only a single frequency and connection approach to modeling large-scale systems origi-
compute Z(~o) exactly. This form of the problem was first nated over three decades ago, it is likely that its full potential
studied by Saeks (1970) and then extended by Saeks and has not been realized. There are several areas of research in
colleagues (Ransom and Saeks, 1975; Saeks et al., 1972). Its which the component connection approach might lead to new
solution is based on the observation that the connection func- results. One of these is the area of optimal decentralized
tion f can be decomposed into two functions (under the control of large-scale systems.
assumption that L22 = 0) as:

f=hog. (7.12)
References
Bello, P.A. et al. (1973). Nonlinear system modeling and analysis.
Report RADC-TR-73-178. Rome Air Development Center, AFSC,
Here, g is a nonlinear function that maps the component
Griffiss AFB, New York.
parameter matrix Z into an intermediary matrix R via:
Chang, R.R. (1973). System identification from input/output data.
M.S. Thesis, Department of Electrical Engineering, University of
R = g(Z) = (1 - Z L l l ) - l Z , (7.13) Notre Dame.
Cooper, G.R., and McGillem, C.D. (1971). Probabilistic methods of
and h is a linear function that maps R into the external signal and system analysis. New York: Holt, Rinehart, and Winston.
parameter matrix S via: DeCarlo, R.A. (1983). Sensitivity calculations using the component
connection model. International Journal of Circuit Theory and
S = h(R) = L21RL12. (7.14) Applications 12(3), 288-291.
DeCarlo, R.A., and Rapisarda, L. (1988). Fault diagnosis under A
The left inverse o f f is given in terms of g and h via: limited-fault assumption and limited test-point availability. IEEE
Transactions on Circuits Systems and Signal Processing 4(4),
f-L = g-%h-L. (7.15) 481-509.
DeCarlo, R.A., and Saeks, R.E. (1981). Interconnected dynamical
systems. New York: Marcel Dekker.
A little algebra will reveal that g-L always exists and is given by Garzia, R.E (1971). Fault isolation computer methods. NASA
the formula: Technical Report CR-1758. Marshall Space Flight Center, Hunts-
ville.
Z = g-L(R) = (1 + RLH)-LR. (7.16) Lee, Y.W., and Schetzen, M. (1961). Quarterly progress report 60.
Research Laboratory of Electronics. Cambridge: MIT.
Thus, the fault isolation problem is reduced to that of finding Liberty, S.R., and Saeks, R.E. (1973). The component connection
the left inverse of the linear function h. This is most easily done model in system identification, analysis, and design. Proceedings of
by working with the matrix representation of h as: the Joint EMP Technical Meeting. 61-76.
Liberty, S.R., and Saeks, R.E. (1974). The component connection
model in circuit and system theory. Proceedings of the European
h = [LT @ L21], (7.17)
Conference on Circuit Theory and Design. 141-146.
Lu, F., and Liu, R.W. (1972). Stability of large-scale dynamical systems.
where ® is the matrix tensor (Kronecker) product (Ransom,
Technical Memorandum EE-7201, University of Notre Dame.
1972; Saeks et al., 1972; Rao and Mitra, 1971). Standard matrix Mattson, S.E., Elmqvist, H., and Otter, M. (1998). Physical system
algebraic techniques can be used to compute h -L. Indeed, the modeling with modelica. Journal of Control Engineering Practice 6,
rows of h can each be identified with external input/output 501-510.
"gains." Thus, the problem of choosing a minimal set of exter- Prasad, N.S., and Reiss, J. (1970). The digital simulation of intercon-
nal parameters for fault analysis is reduced to the problem of nected systems. Proceedings of the Conference of the International
choosing a minimal set of rows in h that render its columns Association of Cybernetics.
linearly independent (Singh, 1972). There are difficulties with Ransom, M.N. (1972). A functional approach to large-scale dynam-
this technique in certain cases, but fortunately, a number of ical systems. Proceedings of the lOth Allerton Conference on Circuits
approaches have been developed for alleviating these (Ransom, and Systems. 48-55.
Ransom, M.N. (1972). On-state equations of RLC large-scale dynam-
1973, 1972; Ransom and Saeks, 1973b).
ical systems. Technical Memorandum EE-7214, University of Notre
Dame.
Ransom, M.N. (1973). A functional approach to the connection of a
7.6 Concluding Remarks large-scale dynamical systems. Ph.D. Thesis, University of Notre
Dame.
Although this chapter has not provided a complete or detailed Ransom, M.N., and Saeks, R.E. (1973a). Fault isolation with insuffi-
explanation of component connection philosophy, the reader cient measurements. IEEE Transactions on Circuit Theory CT-20(4),
should now have a fundamental understanding of the power 416-417.
1084 Stanley R. Liberty

Ransom, M.N., and Saeks, R.E. (1973b). Fault isolation via term mization. Proceedings of the Third Asilomar Conference on Circuit
expansion. Proceedings of the Third Pittsburgh Symposium on Mod- and Systems. 665-670.
eling and Simulation. 224-228. Saeks, R.E. (1969). Studies in system simulation. NASA Technical
Ransom, M.N., and Saeks, R.E. (1975). The connection function- Memorandum 53868. George Marshall Space Flight Center, Hunts-
theory and application. IEEE Transactions on Circuit Theory and ville.
Applications 3, 5-21. Saeks, R.E. (1970). Fault isolation, component decoupling, and the
Rao, C.R., and Mitra, S.K. (1971). Generalized inverse matrices and its connection groupoid. NASA-ASEE Summer Faculty Fellowship Pro-
applications. New York: John Wiley & Sons. gram Research Reports. 505-534. Auburn University.
Rapisarda, L., and DeCarlo, R.A. (1983). Analog mulitifrequency fault Saeks, R.E., Singh, S.P., and Liu, R.W. (1972). Fault isolation
diagnosis. 1EEE Transactions on Circuits and Systems CAS-30(4), via components simulation. IEEE Transactions on Circuit Theory
223-234. CT-19, 634-660.
Reisig, D., and DeCarlo, R.A. (1987). A method of analog-digital Singh, S.P. (1972). Structural properties of large-scale dynamical
multiple fault diagnosis. IEEE Transactions on Circuit Theory and systems. Ph.D. Thesis, University of Notre Dame.
Applications 15, 1-22. Singh, S.E, and Trauboth, H. (1973). MARSYAS. 1EEE Circuits and
Richardson, M.H. (1969). Optimization of large-scale discrete-time Systems Society Newsletter Z
systems by a component problem method. Ph.D. Thesis, University Trauboth, H., and Prasad, N.S. (1970). MARSYAS---Asoftware system
of Notre Dame. for the digital simulation of physical systems. Proceedings of the
Richardson, M.H., Leake, R.J., and Saeks, R.E. (1969). A component Spring Joint Computing Conference. 223-235.
connection formulation for large-scale discrete-time system opti-

S-ar putea să vă placă și