See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3023005

Differential Petri nets: Representing continuous systems in a discrete-event

world

Article  in  IEEE Transactions on Automatic Control · May 1998

DOI: 10.1109/9.665073 · Source: IEEE Xplore

CITATIONS READS
133 846

2 authors, including:

Isabel Demongodin
Aix-Marseille Université
66 PUBLICATIONS   684 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

CPSecurity View project

Batches Petri nets View project

All content following this page was uploaded by Isabel Demongodin on 11 September 2013.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998 573

[9] G. J. Pappas and S. Sastry, “Toward continuous abstractions of dy- In this initiatory phase of hybrid control systems theory, most of the
namical and control systems,” Univ. California, Berkeley, Tech. Rep. discussions focus on the issue of suitable modeling approaches to the
UCB/ERL M96/53, Oct. 1996.
representation of hybrid systems. Different approaches of modeling
[10] J. Raisch, “Control of continuous plants by symbolic output feedback,”
in P. Antsaklis, W. Kohn, A. Nerode, and S. Sastry, Eds., Hybrid have been used, and there is already an abundance of models [16],
Systems II, Lecture Notes in Computer Science, vol. 999. New York: [7]. Some authors (e.g., [4]) define a homogeneous model which links
Springer-Verlag, 1995, pp. 370–390. the discrete-event part and the continuous part in a single formalism.
[11] J. Raisch and S. D. O’Young, “A DES approach to control of hybrid Others (e.g., [8], [1], and [17]) use specific formalisms for each of the
dynamical systems,” in Hybrid Systems III, Lecture Notes in Computer
Science, vol. 1066, R. Alur, T. A. Henzinger, and E. D. Sontag, Eds. two parts or define a model based on the interface between the two
New York: Springer-Verlag, 1996, pp. 563–574. parts. Perhaps naturally, most of the efforts are based on the discrete-
[12] , “A totally ordered set of discrete abstractions for a given hybrid or event part and involve models for discrete-event dynamic systems
continuous system,” in Hybrid Systems IV, Lecture Notes in Computer (DEDS’s) such as finite state machines, process algebras, Petri nets,
Science, vol. 1273, P. Antsaklis, W. Kohn, A. Nerode, and S. Sastry,
Eds. New York: Springer, 1997, pp. 342–360. temporal logic, etc. A nice overview, along with a first attempt for
[13] P. J. Ramadge and W. M. Wonham, “Supervisory control of a class of a unified model, appeared in [7].
discrete event systems,” SIAM J. Contr. Optimiz., vol. 25, pp. 206–230 Among the DEDS models, Petri nets proved to be a very popular
1987. model with the academic and industrial community alike. Having
[14] J. C. Willems, “Paradigms and puzzles in the theory of dynamical
systems,” IEEE Trans. Automat. Contr., vol. 36, pp. 259–294, 1991. a dual nature of a graphical tool and a mathematical object can
serve in both the practical and the theoretical camp. There is a
constantly growing number of publications regarding Petri nets and
their applications which can be found in quite diverse fields. A
major step in the effort to enlarge the modeling power of Petri
nets has been their extension known as continuous Petri nets. The
Differential Petri Nets: Representing Continuous motivation was the inability for successful modeling of discrete-
Systems in a Discrete-Event World event systems with a large number of mostly unobservable events
(e.g., circulation of bottles in a bottling line). Thus, a continuous
Isabel Demongodin and Nick T. Koussoulas approximation was proposed instead, replacing the uncertain counting
with a speed approximation, a technique quite popular in similar
Abstract— Differential Petri nets are a new extension of Petri nets. settings, such as queuing networks. Continuous Petri nets are thus
Through the introduction of the differential place, the differential tran- approximations to discrete-event systems allowing faster simula-
sition, and suitable evolution rules, it is possible to model concurrently tion of the latter without sacrificing accuracy. The combination of
discrete-event processes and continuous-time dynamic processes, repre- continuous with ordinary Petri nets leads to the concept of hybrid
sented by systems of linear ordinary differential equations. This model
can contribute to the performance analysis and design of industrial
Petri nets [2], where hybridity does not refer to the kind indicated
supervisory control systems and of hybrid control systems in general. above.
It is advantageous, if not indispensable, to be able to represent
Index Terms—Differential Petri nets, discrete-event dynamic systems,
hybrid systems, Petri nets, supervisory control systems. both continuous and discrete parts of a hybrid system in the same
context. This can be problematic, however, due to the mathematical
incompatibility between instantaneous events and the convenience
I. INTRODUCTION of continuity. Given that, it appears less cumbersome to choose
One of the most recent and most intense efforts in control theory the discrete-event domain as the environment for this common
deals with handling dynamic systems that include not only the representation. Therefore, an extension for Petri nets is necessary
technological process but its supervisory mechanism(s) as well. For to allow them to represent the continuous time dynamic components.
our purposes, hybrid systems are considered to be all combinations Bourjij et al. [6] show that it is possible to model a hybrid Petri
of a continuous plant (such as a chemical process) or a mixed net in a singular system, by using Euler approximation. Their
continuous/discrete-event plant (such as a chemical process with approach permits us to perform diagnostics through a reference model
process-related logic), with a discrete-event supervisor that reacts to established by an extension of a hybrid Petri net. However, this
external events (planned or unforeseen). Thus, a supervisory control consideration does not take into account the possibility of negative
system of the classical hierarchically structured form of three levels values for continuous variables. To represent negative values, Saadi
(execution, supervision, and coordination) falls into this description. et al. [19] have developed another extension of continuous Petri nets
This kind of control system, being a mixture of continuous-time in which a continuous place can support a negative marking. They
and discrete-event dynamic processes, has been termed “hybrid” or call this extension a dynamical continuous Petri Net and merge this
“discontinuous.” Modeling, analysis, control, and synthesis of such kind of Petri net with a regular one as defined in the hybrid Petri net
systems pose a number of challenging problems. framework. In [5], the simulation of ordinary differential equations is
represented with predicate/transition Petri nets. While they consider
Manuscript received September 23, 1997. This work was supported in part
by the European Commission through the ESPRIT-8924 program SESDIP. only Euler integration because their work seems to be focused on
I. Demongodin is with the Department of Automatic Control and Production real-time applications, it is in principle valid for other integration
Systems, Ecole des Mines de Nantes, La Chantrerie, B.P. 20722, 44 307 algorithms too. In a similar vein, in [15], to arrive at a unified model
Nantes, Cedex 03, France. for the hybrid system, a Petri net equivalent to the causal graph of
N. T. Koussoulas is with the Laboratory for Automation and Robotics,
Electrical and Computer Engineering Department, University of Patras, 26500
the continuous system is found. However, the issues of interface with
Rio, Patras, Greece (e-mail: [email protected]). the “logical” part and the issue of negative markings are not covered
Publisher Item Identifier S 0018-9286(98)02780-9. in both works.

0018–9286/98$10.00  1998 IEEE

574 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998

Fig. 1. Nodes of a DPN.

Fig. 2. Explicit representation of a discrete implicit differential transition.

In this work, the fundamental contribution is the unified Petri to model a hybrid system where the continuous and the discrete parts
net-based representation of both parts of a hybrid dynamic system. are influencing each other’s evolution. A further difference with the
Evidently, the main effort is to represent the continuous part in terms model presented here is that the arc weights can only be positive.
of Petri nets. To make design issues and procedures more transparent, The above may create difficulties in the representation of common
we tried to deviate as little as possible from the concepts and tenets of industrial controller functionalities.
ordinary Petri nets. Thus, we created our extension by making use of Having finished this brief review of various Petri nets related to the
the principles established in the context of continuous Petri nets with representation of continuous quantities, we define now a new class of
maximal speeds depending on time [2], of hybrid Petri nets [9], and Petri nets, which we call DPN. They possess the advantages of the
of dynamic Petri nets [19]. In our approach, the so-called “invariant continuous Petri nets with maximal speed depending on time, those
phases” are equivalent to the integration step that would be used to of the DHPN’s and of course those of the regular Petri nets. Under
numerically integrate the differential equations. The result is a new the assumption that the continuous system can be represented by a
type of Petri net, which we call a differential Petri net (DPN), that finite number of linear first-order differential state equations, they
is able to represent the continuous system part and the discrete-event are powerful enough to model a hybrid system in a single graph. A
system part of a hybrid system in a collective model. DPN is composed of: 1) discrete places and discrete transitions, just
DPN’s are fully defined in Section II. In Section III, we present the as in ordinary Petri nets and 2) differential places and differential
evolution rules for DPN’s, while Section IV discusses their general transitions (symbolism in Fig. 1). To accommodate particular model-
properties and behavior. Section V contains an example along with a ing needs, it is possible to include in DPN’s continuous places and
discussion of capabilities and issues related to DPN’s and is followed continuous transitions [9].
by the conclusion. In the same way as that in a DHPN, the marking of a differential
place is a positive, negative, or null real, representing a state variable
II. DIFFERENTIAL PETRI NETS of the continuous system that is modeled. To every differential
transition we associate a firing speed representing either a variable
A first attempt to break through the barrier of representing only
proportional to a state variable (or a marking of a differential place)
integer quantities within the Petri nets framework has been the
or an independent variable. Since a differential transition is always
continuous Petri net model [9]. There have been a number of
enabled, to discretize the continuous system, we introduce to every
continuous Petri net versions: 1) the constant speed; 2) the variable
differential transition a firing frequency representing the integration
speed; and 3) the asymptotic continuous Petri nets. Their differences
step that would be used when carrying out an integration of the
lie in the way the stream of events is approximated [3], [20].
differential equation. According to the Petri net theory, this delay
In the case of the constant-speed continuous Petri nets (CCPN),
is associated to the implicit discrete transition linked by a discrete
each transition is associated with a constant maximal speed which
place to this differential transition as shown in Fig. 2.
cannot be changed; however, to represent, for example, some control
The intrinsic characteristics associated to a differential transition do
objective, the maximal firing speed must depend on time. In this
not allow the definition of autonomous DPN’s. In fact, inherent in the
way, two extensions of CCPN have been defined, one in [13] called
differential transition is the notion of time, permitting the discretized
continuous Petri net with maximal speeds depending on time and
“view” of continuous systems with a certain period (namely, the
another in [14] named controlled continuous Petri net. In these two
integration step). Thus, a DPN is, by definition, a timed differential
models, the time function associated with continuous transitions can
Petri net.
be varied.
The above nets allowed only positive or null markings. Instead, the
dynamic hybrid Petri net (DHPN) [19], which is a combination of a A. Timed Differential Petri Nets—Definition of Structure
dynamic continuous Petri net with a discrete Petri net, has markings
h Ji
Definition 1: A timed differential Petri net is defined by B =

that can be positive or negative. DHPN has an incidence matrix that R; f; M0 ; verifying the following conditions.
is in block diagonal form, facilitating the search for invariants. This 1) R is a Petri net defined by R = P; T ; h Pre Posti
; with

: finite set of places with j j 1

Petri net extension permits us to model in a unified representation
• P
: finite set of transitions with j j 1
a hybrid system whose continuous part simply cooperates with a P P = n < ;

discrete-event part (as a supervisor). However, it does not allow us • T T = m T < ;

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998 575

• P \ = ; and [ 6= ;;
T P T

• Pre( Pi ; Tj ) is a function that defines arcs from a place

to a transition;
• Post( ) is a function that defines arcs from a tran-
Pi ; Tj

sition to a place.
2) f: P[ T ! fD; DFg, named “differential function,” indicates
for every node whether it is a discrete or differential node.
3) M0 is the initial marking.
4) J is a map, to be called “timing map,” that associates a real
Fig. 3.

number to every transition that can evolve in time and also a

delay for each differential transition (representing the step of The DPN as defined above is a bipartite graph. Places and
integration). transitions are linked by oriented arcs. The associated values of arcs
By definition and for all the formalization, we call a discrete implicit from a place to a transition are those of the Pre function, while
differential transition simply a differential transition. Regarding the the associated values of arcs from a transition to a place are those
Pre and Post functions of a DPN we have: of the Post function. The above scheme can represent in principle
• For a discrete-place: If f (Pi ) = D, then (Pi ; Tj ) and Pre any form of discrete approximation of an ordinary differential equa-
Post
(Pi ; Tj ) are positive integers.
tion. Besides the simple Euler approximation used here, all ODE
• For a differential place: If f (Pi ) = DF, then (Pi ; Tj ) and Pre integration schemes, such as Runge–Kutta, Bulirsch–Stoer, etc., can
Post
(Pi ; Tj ) are real numbers.
be represented within the above framework. Variable-step methods,
where some type of logic is used to decide on the magnitude of each
Because of the presence of negative weights in the net, some
step, can perhaps serve as a more striking example of modeling via
conditions must be defined to preserve the coherence of the model.
DPN.
More precisely, a negative weight can be associated to the arc that
links a discrete or a differential transition to a differential place.
Denoting by mi (t) the marking of a place Pi at time t, then its B. Particular Structure Condition of a Differential Petri Net
value depends on the kind of place: Due to the discretization of the continuous system through a Petri
• For a discrete place the marking is a nonnegative integer: If net formulation, there is a condition between places and transitions
f (Pi ) = D, then m(Pi ) 2 : that must hold. This is a most important element in the modeling
• For a differential place the marking is a real number: If f (Pi ) = process since it introduces the integration step. The following notation
DF, then m(Pi ) 2 : will be used:  Tj ; Tj are the sets of predecessors and successors,
Regarding the timing map of a DPN the following hold: respectively, of the transition Tj ; and  Pi ; Pi are the sets of
• For a discrete-transition: If f (Tj ) = D, then J (Tj ) = dj = predecessors and successors, respectively, of the place Pi .
delay associated at the discrete transition Tj , expressed in time 1) Condition: If an arc joins a discrete place to a differential
units; transition, there must exist a reciprocal arc linking this differential
• For a differential transition: If f (Tj ) = DF, then J (Tj ) = transition to the same discrete place. Further, these two arcs must
hV (Tj ) j dii = hVj j hi; have the same weight, i.e.,
where Vj is the maximum firing speed associated with the transition
Tj , expressed in parts per time unit, and h is the time constant
if f (Pi ) = D and f (Tj ) = DF
(actually, the step in the integration of the differential equation)
associated with the discrete transition linked by a discrete place
such that Pi 2 Tj (resp. Pi 2 
Tj ),

to this differential transition (see Section II-B below). The maximal then Pi 2 Tj

(resp. Pi2 
Tj )

firing speed may be a constant, a linear combination, or a nonlinear and Pre( Pi ; Tj ) = Post( Pi ; Tj ):

function of the markings of the differential places connected to the

differential transition
n This condition assumes that the marking of all discrete places is a
Vj = c; or Vj = ai mi nonnegative integer (because pre- and post-relations for a discrete
i=1 place are positive only), regardless of the evolution of the net. Fig. 3
or shows the above in pictorial format.
Vj = g (m1 ; m2 ; 111 ; mn )

where mi = m(Pi ); Pi 2 P; f (Pi ) = DF: C. Marking of a Timed Differential Petri Net

The maximal speed cannot be a function of derivatives or integrals The state of a Petri net is defined by its marking. For a timed
of the markings since only the instantaneous values are available at differential Petri net, the notion of time associated to the discrete
a given time. As each marking evolves in time, this maximal firing transition introduces a decomposition of the marking. In fact, when
speed changes at the end of each interevent period (similarly to the a discrete transition is enabled at a date t, the tokens required to
case of DHPN). To every differential transition, an instantaneous fire this transition are reserved during the associated delay. When
firing speed vj (t) is associated. This instantaneous firing speed is the delay is over, the transition is fired and the tokens (or marking)
equal to the maximum firing speed or is equal to zero (if the transition reserved for firing are removed from the preplaces of this discrete
is not enabled—see Section III below). Thus transition, while the nonreserved tokens (or marks) are added to the
if f (Tj ) =DF: 0 1  vj (t)  +1
postplaces. Also, at any time, the present marking m is the sum of
two markings, mr and mn , where mr is the reserved marking and
and m
n
is the nonreserved marking. We denote by mi (t) the marking of
vj (t) =V (Tj ) or vj (t) = 0: a place Pi at time t. Then, we have mi (t) = mir (t) + min (t).
576 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998

III. EVOLUTION RULES FOR DIFFERENTIAL PETRI NETS not enabled, the speed vj (t) = 0. If it is enabled, then vj (t) = V (Tj ),
The DPN evolution is realized by the firing of transitions. The depending on the state variables, e.g., vj (t) = i ai mi (t), if a linear
sequence of firings changes the marking of preplaces and postplaces combination is selected.
of the transition. We consider an evolution of this model with a
functioning at maximal speed [9]; in other words, as soon as a discrete IV. STRUCTURE, PROPERTIES, BEHAVIOR AND EVOLUTION
transition is enabled, the tokens or marks required for its firing are Structure-wise, DPN’s are a lot like regular Petri nets. There are,
reserved during all its associated delay. For every type of transition, however, some important differences in the allowable connections
the conditions for enabling and the conditions for firing are different. between elements. The most important feature is that in the connec-
We begin by the conditions for enabling. tion between a differential transition and a discrete place both arcs
Definition 2: A discrete transition is enabled at time t iff each with integer positive weights must necessarily be present (mentioned
preplace has a nonreserved marking at least equal to the weight of earlier in Section II-B). These connections realize one part of the
the corresponding arc interface between the “logic” part and the process.
f (Tj ) = D; Tj is enabled iff Considering conflicts in the strictly differential part of a DPN (i.e.,

8Pi 2  Tj )m i (t)  Pre(Pi ; Tj ):

n
the ODE representation), there may be structural conflict, but there
can be no effective conflict. This is due to the way differential
Definition 3: A differential transition is enabled at time t iff each transitions are enabled and fired. In the modeling of supervisory
pre-discrete place has a marking at least equal to the weight of the systems [11], it is possible to have cases where structural conflicts (in
corresponding arc the sense of ordinary Petri nets) are present, typically a differential
place “feeding” two or more transitions. If one insists in viewing
f (Tj ) = DF; then Tj is enabled iff 8Pi 2  Tj things in terms of ordinary Petri nets, the “conflict” is resolved
such that f (Pi ) = D ) mi (t)  Pre(Pi ; Tj ): beforehand by accepting that the necessary amount of marks is “fed”
to the downstream transitions, as required by the structure of the net.
Note: The enabling is independent of the predifferential places’ The absence of effective conflicts means that the reachability tree
marking, as the latter can be positive or negative, i.e., continuously that corresponds to the differential part will have no branching. On
varying. the contrary, branching may very well be present in the discrete Petri
An enabled discrete transition can be fired. If a timing mapping net part.
is associated to the discrete transition, the fact that the transition is An incidence matrix W can be defined for every DPN
enabled “releases” the time and reserves the relevant marks. At the
end of this time interval, the transition can be fired. When a discrete
W = [Wij ]n 2m ; where Wij = Post(P ; T ) 0 Pre(P ; T ):
i j i j

transition is fired: 1) a number of tokens equal to the weight of arc Clearly, after suitable ordering, this matrix can have a block structure
joining (upstream) places to the discrete transition are removed from where the two diagonal blocks represent the dynamics of the continu-
the preplaces of the discrete transition and 2) a number of tokens equal ous and the discrete-event subsystems, while the off-diagonal blocks
to the weight of the arc joining the discrete transition to (downstream) represent the interface between those two subsystems. Furthermore, if
places are added to the postplaces of the discrete transition, i.e., we consider the characteristic vector  (t) containing a firing sequence
f (Tj ) = D and J (Tj ) = dj for the discrete transitions and a speeds vector v (t) containing the
Tj is fired at time t + dj instantaneous firing speeds of the differential transitions, then it is
 possible to deduce from a given marking at time ti the reachable
) 88PPii 22 TTj : :mmii((tt++ddjj))==mmii((tt))+0 Post
Pre(Pi ; Tj )
(Pi ; Tj )
marking at the date tk , by the following relation (similar to the
j “fundamental equation” in [9])
if no change of mi has occurred between t and (t + dj ). t

An enabled differential transition can be fired. When a differential M (tk ) = M (ti ) + W 1 (tk ) + v(u) du :
t
transition is fired, then: 1) a number of tokens equal to the firing
speed multiplied by the weight of arc joining the (upstream) places Regarding Petri net-type properties, the following can be said.
to the differential transition is removed from the preplaces of the The continuous (sub)system part of a DPN actually represents the
differential transition and 2) a number of tokens equal to the firing workings of an integrator for differential equations. This function is
speed multiplied by the weight of the arc joining the differential naturally approximated as long as the timing h of the differential
transition to (downstream) places is added to the postplaces of the transitions is not zero. As h goes to zero, this part of the DPN
differential transition, i.e., constitutes a better and better approximation of the continuous
dynamic system. This means that if the continuous system is stable,
f (Tj ) = DF and vj (t) firing speed of Tj at the date t; then the markings of the differential places will remain always
Tj is fired at time t during a delay dt ) bounded. When h is not zero, however, whether there will be
boundedness depends only on the numerical behavior of the implicit
• 8Pi 2  Tj : mi (t + dt) = mi (t) 0 vj (t) 1 Pre
(Pi ; Tj ) 1 dt: integration. Thus, a choice for h that would have created numerical
• 8Pi 2 Tj : mi (t + dt) = mi (t) + vj (t) 1 Post
(Pi ; Tj ) 1 dt: instability in the corresponding integration algorithm will force
Remark: The particular condition 1 of a DPN assumes that the unboundedness in the DPN. Liveness is also guaranteed for a DPN
marking of the discrete places is an integer, even if those discrete that represents strictly a continuous dynamic system. However, a
places are upstream places of a differential transition. In fact, the DPN that contains discrete places and transitions may very well have
quantity of marking removed from them by the firing of the differ- deadlocks. By studying the DPN model of a supervisory system, one
ential transition is immediately added back to them. can determine how the overall system will behave in case of faulty
For a differential transition in a DPN, calculation of the instanta- logic that leads to a deadlock situation.
neous firing speed depends on its state, in other words on whether it is A behavior algorithm of a DPN (Fig. 4) is defined in order to
enabled or not. To compute the instantaneous firing speed we use the simulate real systems, e.g., an industrial production process. This
J function of the transition. Then, if the differential transition Tj is behavior algorithm is composed of two essential parts. The first part
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998 577

differential transition (and are usually hidden in the representations)

should have been normally included in the evolution graph. However,
since they do not affect the evolution (there is no information loss),
they can be omitted. The nodes of the evolution graph are linked by
arcs that determine the occurrence of events and the “expended” delay
between two consecutive interevent states. The transition associated
to this arc is denoted by hoccurrence of events j dk i, where dk is the
elapsed time. The evolution graph may have branching which will
be the result of effective conflicts in the discrete part of the DPN.
The graph will normally be of infinite length unless there is a finite
horizon for our system.

V. EXAMPLE AND DISCUSSION

We consider a system whose behavior is described by an alternation
between the following two models:
1) x_ (t) = 06.
Fig. 4. Behavior algorithm of a DPN. 2) x_ (t) = 0x(t).
The switching between these two models is governed by a simple
rule. For two time units, the system admits model 1), while for the
calculates the enabled transitions and the firing speed of differential following three time units model 2) is valid; then the same cycle
transitions, whereas the second part determines the nearest event date. repeats. At the switching instant, there occurs a jump of constant and
All events that change the state of the system are computed, the known magnitude in the state of the system. Therefore, the rules are:
1) for kT  t < kT + 2T =5; k = 0; 1; 1 1 1 use model x_ (t) = 06;
nearest event being the nearest in time. The date of this event becomes
2) for kT + 2T =5  t  (k + 1)T ; k = 0; 1; 1 1 1 set x(t)
the current date, and at this instant, the new place markings are
_ (t) = 0x(t):
determined. At a fixed date and for a given marking, the identification
x(t) + 4 and use model x
of enabled transitions allows one to calculate the firing speeds
associated to the differential transitions, defined in the previous We choose the integration step h = 0:5, which is crude but effective
paragraphs. for demonstration. The DPN model appears in Fig. 6. Note that both
The determination of the next event dates depends on the evolution the switching times and the jump could have arbitrary values, as
of the state of places, according to which transitions are enabled. long as their generating mechanism can be modeled in DPN terms.
Between two event dates, the state of the system remains constant. We defined the following sets:
This particular behavior and the explanation of what is meant by
“state” are described in the following definition.
PD f 1 2 g discrete places
= P ;P ;

Definition 4: The interevent state of a DPN corresponds to a TD f 1 2 g discrete transitions

= T ;T ;

period of time during which the elements of the state of the net, PDI = f 3 4 g implicit discrete places
P ;P ;

DI = f 3 4 g discrete implicit transitions

comprised by: 1) the markings in the discrete places; 2) the instanta- T T ;T ;

DF = f 5 g differential places
neous firing speeds of the differential transitions; and 3) the reserved
P P ;
markings of discrete and differential places, remain constant.
From this definition, to compute the next event dates, it is essential TDF = f 5 6 g differential transitions.
T ;T ;

to know the types of events which can change an interevent state.

Property: The interevent state of a timed differential Petri net Then, the Pre-, Post-, and incidence matrices are
changes if and only if one (or possibly several at the same time) 0 1 0 0 1 0
of the following kinds of events occurs. 1 0 0 0 0 1
1) firing of a discrete transition;
2) the nonreserved marking of a place which is a preplace of Pre = 0 0 1 0 1 0
a discrete transition becomes equal to the weight of the 0 0 0 1 0 1
corresponding arc.
The evolution graph of a DPN may be represented by nodes corre- 0 0 0 0 0 0
sponding to interevent states and transitions between these nodes. A 1 0 0 0 1 0
node is divided in two parts. 0 1 0 0 0 1
1) The first part represents the markings of the discrete places.
MD = fMi such that f (Pi ) = Dg is the total marking of the Post = 0 0 1 0 1 0
discrete places and MDr = fMir such that f (Pi ) = Dg is their 0 0 0 1 0 1
reserved marking.
2) The second part represents the firing speed vector of the 4 0 0 0 01 06
differential transitions and the reserved markings MDF r
= 1 01 0 0 0 0
fMir such that f (Pi ) = DFg. The total markings of the 01 1 0 0 0 0
differential places MDF = fMi such that f (Pi ) = DFg at
the beginning and at the end of the phase are also included. W = 0 0 0 0 0 0 :

Fig. 5 shows this representation and all the information of the 0 0 0 0 0 0

evolution graph’s node. It is to be noted that the implicit discrete
places and discrete timed transitions that are connected to each 4 0 0 0 01 06
578 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998

Fig. 5. Node of the evolution graph.

Fig. 6. DPN representation of the system.

The evolution graph for this DPN and further details can be found and analyzed through DPN’s. Just as in other hybrid systems models,
in [10] and [12]. complexity is a cause for concern, while integrated design cannot
The above example shows that DPN can be used in the modeling be supported at this time. Although it is possible to use linear and
of hybrid systems. At this point it is possible to model linear multi- robust control theory techniques for controller design, and heuristics
variable systems under a first-order approximation. It is possible to or newly proposed design methods for Petri nets (e.g. [21]), it is not
extend DPN to cover systems that are more complex. It is anticipated, possible to do both concurrently. Such issues will be the topics of
however, that the simplicity of the representation may be lost and the further investigations.
deviation from classical Petri nets may become greater. Nonetheless,
DPN can contribute to the analysis of one of the archetypal hybrid VI. CONCLUSION
control system, i.e., the industrial supervisory control system. Such an The DPN constitutes a new modeling tool suitable for systems
effort was the subject of [11] (full details in [12]). It was shown there having a combination of discrete-event evolution and continuous
how one can model typical functionalities of industrial controllers, state evolution, such as supervisory systems widely used in industrial
such as level crossing, bound violations, testing variation within a automation. This new extension of Petri nets has many common
region of values, etc. The increasing acceptance of Petri nets as a characteristics with ordinary Petri nets and can incorporate extensions
vehicle for software and protocol modeling implies that distributed of ordinary Petri nets, such as stochastic, interpreted, etc. So far, only
real-time control systems (comprising network communications, logic the representation of linear time-invariant dynamic systems has been
controllers, instrumentation and actuation systems) can be represented covered, but the representation of systems that are more complex
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 4, APRIL 1998 579

appears possible. The contribution of the DPN model to the study of Compositional and Uniform Modeling of Hybrid Systems
hybrid control systems can be multifaceted. Analysis is facilitated,
since issues such as deadlock and safeness can be investigated with Albert Benveniste
the behavior of the continuous process being taken into account.
The performance of supervisors, especially of complex ones, can be
assessed more easily. Finally, the performance of integrated industrial Abstract—This paper discusses fundamentals of hybrid system model-
ing. Emphasis is put on compositionality and the use of multiform time.
automation systems can be evaluated.
Compositionality refers to the ability of freely composing hybrid systems.
Since hybrid systems are considered, different time indexes occur, and
modularity calls for considering “time” as local to each module, this is
REFERENCES what is called multiform time. The proposed framework is behavioral,
hence a way is provided to automatically synthesize proper scheduling
[1] P. Antsaklis, J. Stiver, and M. Lemmon, “Hybrid systems modeling and constraints for the joint simulation of discrete-time, event-based, and
autonomous control systems,” R. L. Grossman et al., Hybrid Systems, continuous-time components. Finally, the relations of this model with the
Lecture Notes Computer Science, vol. 736. New York: Springer, more traditional state-based point of view are discussed.
1993.
[2] J. Le Bail, H. Alla, and R. David, “Hybrid Petri nets,” in Proc. 1st Index Terms—Concurrency, hybrid systems.
European Control Conf., Grenoble, France, July 1991, pp. 187–191.
[3] , “Asymptotic continuous Petri nets: An efficient approximation of
discrete event systems,” in Proc. IEEE Robotics and Automation Conf., I. INTRODUCTION
Nice, France, 1992. Hybrid systems have been a topic of growing interest and ac-
[4] A. Benveniste and P. Le Guernic, “Hybrid dynamical systems theory
and the SIGNAL language,” IEEE Trans. Automat. Contr., vol. 35, pp. tivity in the recent years. It is our opinion that relevant real-life
535–546, May 1990. applications of hybrid systems paradigm mainly consist of very
[5] M. Brielmann and B. Kleinjohann, “A design model for concurrent large complex distributed systems. For such systems, modularity in
engineering of heterogeneous systems,” in Proc. EURODAC 1993, pp. modeling, simulation, control design, and verification is mandatory.
158–163.
[6] A. Bourjij, M. Zasadzinski, M. Darouach, and J. L. Preudhomme,
Thus, we shall request for our model that: 1) composition works
“Diagnostic par les réseaux de Petri hybrides: Application a un réseau de uniformly for both discrete and continuous parts, and is based on
générateurs de vapeur simule,” in Proc. Automation of Mixed Process: easy and clean mathematical principles and 2) “time” shall be local
Dynamical Hybrid Systems, Symp. AFCET/ADPM’94, Bruxelles, Bel- to each module instead of global. Point 1 calls for a composition
gique, pp. 239–246. via the intersection of behaviors, or, equivalently, via conjunction of
[7] M. S. Branicky, V. S. Borkar, and S. K. Mitter, “A unified framework
for hybrid control,” in Proc. 33rd Conf. Decision and Control, Lake constraints or systems of equations. Point 2 leads us to consider that
Buena Vista, FL, Dec. 1994, pp. 4228–4234. “time bases” should not be given once for all, but rather should be
[8] R. Brockett, “Language driven hybrid systems,” in Proc. 33rd Conf. variables of our hybrid systems. A more extensive introduction and
Decision and Control, Lake Buena Vista, FL, Dec. 1994, pp. 4210–4214. motivations can be found in [1].
[9] R. David and H. Alla, Petri Nets & Grafcet—Tools for Modeling
Discrete-Event Systems. Englewood Cliffs, NJ: Prentice Hall, 1992.
[10] I. Demongodin and N. T. Koussoulas, “Modeling dynamic systems II. A GENERAL HYBRID SYSTEMS MODEL
through Petri nets,” in Proc. IEEE CESA’96 IMACS Symp. Discrete
Events and Manufacturing Systems, Lille, France, pp. 279–284.
[11] , “Modeling of hybrid control systems via differential Petri nets,” in A. Primitives
Proc. 4th IEEE Medit. Symp. Control and Automation, Maleme, Greece,
R
Time: Our time index set is = (01; +1); viewed as a totally
R
June 1996, pp. 428–433.
[12] , ESPRIT-8924 SESDIP Task A.2.3.2 Final Rep., June 1996. ordered dense and continuous set, we also write = [01; +1].
[13] E. Dubois, H. Alla, and R. David, “Continuous Petri net with maximal This choice only specifies that we work with a continuous, totally
speeds depending on time,” in Proc. 15th Int. Conf. Application and
ordered, and one-dimensional time. It does not imply, however, that
Theory of Petri Nets, Zaragoza, Spain, June 1994.
[14] M.-K. Ghabri and P. Ladet, “Controlled continuous Petri nets,” in Proc. we are bound to a single time t, as we shall see later.
IEEE Conf. Robotics and Automation, 1994. Presences: “Time bases” will be used as index sets for signals
[15] M. M. Gomaa and S. Gentil, “Systèmes dynamiques hybrides: Une or for “task activation” (we do not define formally what we mean
nouvelle approche de modélization pour la supervision,” Analyse et by a task here). A time basis set is assigned to each signal. Time
Supervision des Systèmes Dynamiques Hybrides, J. Etudes, Feb. 1996.
bases will be referred to in the sequel as “presences.” A presence is
[16] R. L. Grossman et al., Hybrid Systems, Lecture Notes Computer Science,
vol. 736. New York: Springer, 1993. any Borel1 subset T of R
such that T \ (01; x) = ; for some
[17] L. Jacot and P. Ladet, “A modeling framework for hybrid systems,” in finite x.2 The family of presences is equipped with the following set
Proc. IEEE SMC 1993 Conf., LeTouquet, France, pp. 692–698. theoretic operations: union, intersection, and set difference. Discrete-
[18] C. S. Pan, M.-K. Ghabri, and P. Ladet, “Dynamic Petri net and
time divergent sequences, as well as intervals, are particular cases
its applications in the production,” in Proc. IFAC Integrated Systems
Engineering, Baden-Baden, Germany, 1994. of presences.
[19] J. Saadi, A. Saad, T. Bennani, and H. Alla, “Les réseaux dynamiques hy-
brides—Un outil graphique de représentation des systèmes dynamiques Manuscript received September 23, 1997. This work was supported in part
hybrides,” in Proc. Automation of Mixed Processes: Dynamical Hybrid by NSF/ESPRIT under Grant EC-US-043.
Systems, Symp. AFCET/ADPM, Bruxelles, Belgique, Nov. 1994, pp. The author is with IRISA-INRIA, Campus de Beaulieu, 35042 Rennes
211–220. cedex, France (e-mail: [email protected]).
[20] N. Zerhouni and H. Alla, “Dynamic analysis of manufacturing systems Publisher Item Identifier S 0018-9286(98)02910-9.
using continuous Petri nets,” in Proc. IEEE Robotics and Automation 1 The reader not willing to bother with technicalities can discard the “Borel”
Conf., Cincinnati, OH, 1990.
[21] K. Yamalidou, J. Moody, M. Lemmon, and P. Antsaklis, “Feedback assumption. Note, however, that we need a class of subsets of R which is
control of Petri nets based on place invariants,” Automatica, vol. 32, invariant under union, intersection, and set difference.
no. 1, Jan. 1996. 2 This latter condition expresses that signals have some finite birth date.

0018–9286/98$10.00  1998 IEEE

View publication stats