Nof S.Y. Springer Handbook of Automation

Подождите немного. Документ загружается.

Distributed Agent Software for Automation 28.3 Building Agents for the Curing System 475

quickly detected and assessed by the individual agents

that are affected by such variations. Inconsistency be-

tween observed temperatures provides the basis to

suspect, detect, and further isolate the faulty element.

Fault detection and diagnosis are key enabling fea-

tures of the agent infrastructure that drive the processes

of conﬁguration and planning. Dynamic reconﬁgu-

ration may include changing the operating state of

a group of system devices in a coordinated manner as

well as dynamically changing the control for system

elements.

28.2.2 Agent Tool

The technological advantage provided by the agents

is the ability to create agent components in industrial

commercial off-the-shelf PLCs without having to cre-

ate specialized hardware and programming tools. The

advantage of using PLC devices as the primary control

and agent-hosting platform is the reliability of its op-

erating system, hardware components, and packaging,

which are readily available for industrial use.

By embedding theagents directly intothe controller,

it is possible for the agent to interoperate with the con-

trol programs faster. This embodiment eliminates the

potential loss of controllability due to a lost connec-

tion between a controlling device and the supervising

computer. Control programs ensure the hard, real-time

response required by many control applications and an

agent part ensures high-level coordination and ﬂexibil-

ity. This synergy of agent and control inside the PLC

and in combination with smaller PLCs has made it pos-

sible to transform a physical machine into an intelligent

machine by localizing the intelligence.

The changes made to the PLC ﬁrmware include

software extensions to the tasking model of the con-

troller, object instantiation and retention during power

cycles, and a more general communication layer. Great

effort took place initially to make the PLC an agent-

hosting device with the capability to absorb executable

agent code via object downloading. The extensions to

the communication layer covered the generation and

parsing of FIPA-based messages. With this feature, the

control-based agents talk to other agents that comply

with the FIPA speciﬁcations, even those that may have

been implemented using a different infrastructure.

To make the agent system scalable and manage-

able in terms of software for agent development, it was

necessary to create a development environment for the

programming of agent libraries [28.35]. There are ﬁve

main phases of agent development:

1. Template library design

2. Facility editor

3. Control system editor

4. Assignment wizard

5. Control code generator.

In the template library editor, a collectionof compo-

nents to represent the physical system is identiﬁed and

built with control and reasoning parts. These compo-

nents are generic, without referring to speciﬁc instances

or devices. The library of components is then instan-

tiated and arranged in a speciﬁc tree to represent the

hierarchical order and attributes of the components that

will be part of a speciﬁc system. This is where pa-

rameters of controlled devices and locations are added.

The control system editor helps in the selection of the

control and communication network equipment. In ad-

dition, it helps in the assignment of the input and output

(I/O) points.

Once the pieces above are deﬁned, the control and

reasoning components deﬁned within theapplication (in

the facility editor) are assigned to speciﬁc controllers

(PLCs). After the assignment, the executable code is

generated and downloaded to the controllers.

28.3 Building Agents for the Curing System

The purpose of modeling the composite process us-

ing industrial agents is to increase the degree of

survivability, diagnostics, and reconﬁgurability of the

system [28.36]. Survivability of the control system is

very important since it is directly related to safety and

the creation of scrap material.

Curing is a long process and therefore there is

a greater time span during which disruption may oc-

cur. Situations such as power shutdowns and network

discontinuity are of critical importance and these must

be handled in a timely way to sustain high process in-

tegrity. Since the curing process consists of applying

heat and cooling to composite material inside a sealed

autoclave, a reliable system to operate on top of the

temperature sensors is fundamental. The thermocouples

read the temperature of the composite material at differ-

Part C 28.3

476 Part C Automation Design: Theory, Elements, and Methods

Composite

Thermocouple

Autoclave

Fig. 28.3 Composite curing

ent locations. Moreover, since these constitute a group

of logically interconnected sensors, the thermocouples

are treated as a multisensor problem. Figure 28.3 il-

lustrates how a composite material part is populated

with the multisensor array prior to its placement in-

side the autoclave furnace. Thermocouple agents will

be created to represent the thermocouple devices in the

controller. The agents will use their social skills to or-

ganize the array into a smart and highly reconﬁgurable

logical structure.

Given that the process operates in a controlled vac-

uum, it is not possible to stop the process to adjust

a loose thermocouple, for example. Therefore, the au-

todiagnosability of the process is a critical factor that is

made a part of the decision-making roles of the agents.

As a consequence, both the supervisory and control

roles must be able to detect problems and reconﬁgure

the physical system to cope with such eventualities in

order to enable continuous, uninterrupted activity.

Factorssuch asnonlinearities of the process and ma-

terial (exothermal reactions provoking thermal spikes)

must be considered in the modeling of the supervisory

intelligence. The agent technology discussed herein

allows for an incremental implementation of the neces-

sary rules. The agent’s decision power resides in rules,

which are encapsulated in the behaviors. A validation

framework based on simulation helps to create a virtual

model for testing and enhancing the curing control rules

and loops of the agents and control algorithms.

The agent software that is downloaded to the PLC is

associated with a control-level driver (IEC-61131 con-

trol programs). The agent directs the control system

using a lightweight supervisory strategy. Agents do not

perform direct control. The device drivers carry out di-

rect controlupon theequipment. Thus,to avoiddecision

collisions and misinterpretation of the world, the sphere

of inﬂuence of each agent must be deﬁned with respect

to its associated physical device.

The distribution of the agents throughout speciﬁc

sections of the application is a necessary partitioning of

responsibility that needs to take place before program-

ming the rules. In the current case, the application is

made ofa singlemachine (autoclave)with amultisensor

array (thermocouples).

There are three partitioning criteria to model the

system:

1. One-to-one

2. One-to-many

3. Many-to-one.

One-to-one portioning exists in physical devices

that can be isolated and classiﬁed as independent work

units. Typical examples of such a case are a valve,

switch, or breaker. In this case, one agent per physi-

cal device is enough. A one-to-many case appears in

complex systems, where there is still an easy classiﬁca-

tion of the work units. This type of partitioning happens

in equipment grouping situations, whereby one agent

can be associated with a group of related physical de-

vices. A rarer case is many-to-one partitioning,in which

multiple agents are associated with a single or grouped

Curing system

Autoclave

Thermocouples

Fig. 28.4 Agent

classiﬁcation for

curing system

Part C 28.3

Distributed Agent Software for Automation 28.4 Autoclave and Thermocouple Agents 477

physical device. A typical situation of this type oc-

curs when a single agent cannot cope with too many

concurrent processes.Efﬁciency of the decision-making

process is then measured in terms of communication

latencies and decision-making convergence. A system

with too many agents and processes has many more

messages on the network. The network may become

overwhelmed with too many packets and communica-

tion latencies may increase. Since agents collaborate

with their peers through the exchange of information,

they heavily depend on timely arrival of information

during the group decision process. Thus, it may occur

that the agents begin dropping out of the conversation

without concluding a decision process. In this case,

agents would retry communication, causing more mes-

sages and therefore problems in convergence. Thus,

special care must be taken to obtain a good partitioning

of agent functionality.

The composite curing system corresponds to

a three-tier one-to-one system. As shown in Fig.28.4,

the top-level tier has one agent to represent the overall

curing system. The second tier represents the autoclave

machinery that will be part of the curing system. There

is one autoclave agent per autoclave machine. The third

tier represents the thermocouple sensor level. There is

one agent per thermocouple.

28.4 Autoclave and Thermocouple Agents

A bottom-up design (from the simplest part of opera-

tion to the more complex behavior) of the components

takes place to build the curing system. Each component

is associated with a piece of equipment or a partic-

ular process as an agent. Figure 28.5 shows a group

of curing components. Most of the components are

control-oriented drivers. However, autoclave and ther-

mocouple components are agents.

The left-handside ofFig.28.5 shows the component

list. The upper-right part shows the low-level program-

ming blocks of the component. The lower-right part

Fig. 28.5 Component list

shows the list of control parameters of the component,

such as input and output (I/O) variables.

28.4.1 Autoclave Agent

The autoclave agent contains curing artifacts such as

cooler, heater, pressure compressor, and thermocouples.

Thermocouple agents have their own temperature mon-

itoring system and personality. Depending on the stage

of the curing process, an autoclave agent periodically

requests a reorganization of the thermocouples to se-

Part C 28.4

478 Part C Automation Design: Theory, Elements, and Methods

Template behaviorAgent capability Procedures

Fig. 28.6 Agent capabilities, behaviors, and procedures

lect a leading and a lagging thermocouple. The nominal

behaviors given to an autoclave agent consist of the

following:

•

cNetTCTemp: This behavior generates a multicast

message to all dependent thermocouples (TC1–

TC16 in Fig. 28.2 to have them summarize their cur-

rent temperature trends. Asynchronous responses

are sent by the thermocouple agents to report their

temperature information.

•

SetNewLeaTC: In this behavior, the autoclave agent

applies process-speciﬁc criteria to generate a list of

leading and lagging thermocouples. The autoclave

agent uses a min–max selection mechanism. The

process recipe and the stage of curing determine the

temperature boundaries (upper and lower).

•

startCuringProcess: This behavior contains the

startup sequence of the autoclave.

•

stopCuringProcess:This behavior contains theshut-

down sequence of the autoclave.

The aggregation of capabilities into the agent behavior

creates the rules of the agent incrementally. In Fig.28.6

all autoclave agents have the CuringEvent capability,

which implements a set of template behaviors. Each

behavior is associated with a set of procedures.

28.4.2 Thermocouple Agent

A thermocouple agent has two essential behaviors to

monitor the temperature of the process and the sensor

condition and trending using the following behaviors

•

getTemperature: This behavior uses a sampling

algorithm to calculatetemperature average and stan-

dard deviations.This behavior allowsfor continuous

adjustment of temperature measurement to compen-

sate for natural jitter.

•

provideTemperatureTrend: This behavior enables

the detection of temperature variations outside the

nominal ranges. The feasible range allows for

the detection of extreme deviations from nominal

trends. These events are reported to the autoclave

agent as events.

Each thermocouple carries out continuous sam-

pling of the temperature. This information is processed

statistically. Temperature sampling is an internal-to-

thermocouple action, which corresponds to an intera-

gent communication. The thermocouple agents receive

requests from the autoclave agent to transmit their

temperature trends in the form of agent messages in

a periodic fashion.

28.5 Agent-Based Simulation

Simulation is made up of two parts: the machine and the

material. Machine simulation covers the autoclave ma-

chine itself with its mechanical devices (such as heater,

cooler, and compressor) and thermocouples. Material

simulation mimics the behavior of the composite ma-

terial. A ﬁnite-difference mesh simulates the thermal

dissipation throughout the material. The model repre-

sents a single piece of composite material of square

shape with arbitrary sides and thickness.

The control process takes place in a PLC.How-

ever, for the purpose of the simulation, a soft controller

PLC (Rockwell Automation, SoftLogix 5800) with

ﬁrmware extensions to support the agents was used.

The Simulink/Matlab engine was used as the simula-

tion toolset.To mimic the physical process,a simulation

of the material, thermocouples, and material was cre-

ated. Theagents usethe simulationas the process model

instead of the actual machine and material.

In Fig. 28.7 there are horizontally and vertically ar-

ranged gray rectangular boxes. These boxes contain

heat transfer differential equations to calculate the ac-

cumulation and dissipation of heat on a speciﬁc node

of the ﬁnite-difference mesh. The heat transfer model

is subjected to anisotropic properties to mimic the heat

Part C 28.5

Distributed Agent Software for Automation 28.5 Agent-Based Simulation 479

TC01

Goto

TC05

Goto4

TC06

Goto5

TC02

Goto1

TC03

Goto2

out

T1 T2

out

T1 T2

out

T1 T2

out

T1 T2

out

T1 T2

out

T1 T2

Fig. 28.7 Composite material simulation

dissipation rates throughout the material. The chemical

composition of the resin affects the heat transfer, mak-

ing it vary with location and direction of propagation.

The square boxes represent ﬁnite-difference nodes

(concentration nodes) to calculate the average tem-

perature of the material at the speciﬁc location of

measurement. Figure 28.7 shows just a small section

of the ﬁnite-difference mesh. In the actual model, there

are 16 concentration points and 16 thermocouples. The

thermocouple sensors are scattered throughout the ma-

terial to measure the temperature of the composite at

each node. The temperature readings are sent to the

PLC as an input array of temperatures to be further

interpreted by the thermocouple control drivers.

Figure 28.8 shows the simulation blocks that

contain the autoclave and autoclave’s inner atmo-

sphere. The autoclave agent connects to the simu-

lation throughout six inputs parameters (heaterIsOn,

coolerIsOn, coolerIsEnabled, heaterIsEnabled, heatin-

gRate, coolingRate) to control the autoclave. The

maximum and minimum heating and cooling rates are

provided in the curing recipe. Heating and cooling rates

are used as the control variables to control the amount

of heat that is applied to the material. Table 28.1 shows

an example of the command that initiates a heating-up

phase in the simulation.

Control commands drive the simulation in desired

directions to achieve a particular curing proﬁle. All 16

thermocouple readings must be maintained within the

curing envelope (refer to Fig. 28.1 for the proﬁle shape).

Deviations from the curing envelop are regulated by

the cooperative actions of the agents. The agents contin-

uously monitor the health of the thermocouples using

temperature trending and standard deviations. When

a thermocouple fails, the agent that represents that ther-

mocouple decides to dismiss itself from the monitoring

array. This decision is taken autonomously at the lowest

level. The dismissal cascades up the hierarchy, inducing

Part C 28.5

480 Part C Automation Design: Theory, Elements, and Methods

Autoclave atmosphere

Heating/cooling unit

(heaterIsOn)

(coolerIsOn)

(coolerIsEnabled)

(heaterIsEnabled)

(heatingRate)

(coolingRate)

Fail heater = 0

Fail cooler = 0

heaterIsOn

coolerIsOn

coolerIsEnabled

heaterIsEnabled

heatingRate

coolingRate

Fig. 28.8 Autoclave and autoclave atmosphere simulation

a reorganization of the sensor array to adjust to the new

reading set.

If a departing thermocouple was operating as a lead-

ing or lagging thermocouple, the current stage of curing

is at high risk since the autoclave agent loses its win-

dow into the process variable. The reorganization of

the sensor arrays implies a discovery of new leading

and/or lagging thermocouples to reestablish the process

variable. While the agents reorganize the system, the

control must continue its normal activity by maintaining

all temperatures withinthe nominal rangesas prescribed

in the recipe, but this fall-back position can only be sus-

tained for a short period until the new process variable

is reestablished. This reorganization activity is the core

Table 28.1 Autoclave heat-up command

Variable name Value

heaterIsEnabled true

coolerIsEnabled false

coolerIsOn false

heaterIsOn true

heatingRate %.%

coolingRate %.%

beneﬁt of the multiagent system. Agents are intended to

handle these dynamic decisions.

28.6 Composite Curing Results and Recommendations

The composite curing process begins with download-

ing of the agents and device drivers into the PLC.Next,

the process simulation is started and connected to the

soft controller throughout and automation proxy. The

role of the automation proxy is to exchange I/O data be-

tween the controller and the simulation, as well as the

synchronization of their clocks.

There are several aspects of interest that can be ex-

tracted from this technology. First, there is a permanent

learning process associated with the encapsulation of

the behaviors into particular agents, as well as the parti-

tioning of the domain into regions.

28.6.1 Designing the Validation System

The system designer carries out preliminary veriﬁca-

tions to test the agent behaviors. It is natural to see

some chaotic results in the initial iterations. The de-

signer adjusts the agent and control functions to make

the prototype stable and observable using an interactive

process.

The design of the agent is based on tem-

plates [28.35]; for instance, only one thermocouple

agent template is built for all 16 thermocouples.

Likewise, in object-oriented programming, agent pro-

Part C 28.6

Distributed Agent Software for Automation 28.6 Composite Curing Results and Recommendations 481

gramming allows for the creation of agent instances

from a single agent template. Thus, all thermocouple

agents share the same reasoning rules but vary in per-

sonality (name, location, ranges, etc.).

Another consideration is the speed of the simula-

tion. The real curing process takes hours to complete.

However, during the modeling and validation phase

it is not affordable to spend so much time waiting

for events and changes in the process, so the sim-

ulation has to be accelerated to compress the total

time. This requirement adds complexity to the model

since the controller cycles must be in real time to

mimic actual control. Time compression is factored

into the control timers to make events observable in

shorter time spans. In this curing model, ideal time

spans during the validation phase d should not exceed

30min/cycle. Once the overall model is ready, tests

begin to check the control/agent responses against the

system dynamics.

0 1000 2000 3000 4000 5000 6000 7000

Temperature

Time

Agent curing control–enhanced rules

Heating Cooling

300

250

200

150

100

0 102030405060708090100

100

0 102030405060708090100

Fig. 28.9 Heating and cooling simulation

28.6.2 Modeling Process Dynamics

A critical requirement in agent-based engineering is

the interdisciplinary attitude of the engineers. The cre-

ation of these sophisticated models requires more than

one design perspective to create valid results. The

designer transits from sole software engineer into a con-

trol engineer who understands the dynamics of the

process.

Figure 28.9 shows the distribution of temperature

for one of the test runs. The colored regions represent

isothermals for heating and cooling. The isothermals

change in shape and size as the curing process evolves.

The curing proﬁle is shown in the lower half of

Fig.28.9. The temperature distribution is not uniform

due to the properties of the material. The thermal reac-

tion of the composite makes hot and cold spots change

location and size over time. The control system must

maintain all temperatures within the upper and lower

Part C 28.6

482 Part C Automation Design: Theory, Elements, and Methods

Curing model

Perturbations

PID

Autoclave

agent

Thermocouple

agents

representative

set

Fig. 28.10 Agents and control cooperation

bound of the curing envelope. The proﬁle introduces

abrupt changes in the temperature slopes when mov-

ing from a neutral stage into a heating or cooling stage.

The control system must regulate the overshooting am-

plitude while maintaining a good conﬁguration of the

thermocouple. There is overshooting in the process but

the recipe allows for some operation outside the bound-

aries for short periods.

0 500 1000 1500 2000 2500 3000 3500

a) Temperature (°C)

Time (min)

Agent curing control-simple rules

300

250

200

150

100

0 1000 2000 3000 4000 5000 6000 7000

b) Temperature (°C)

Time (min)

Agent curing control–enhanced rules

300

250

200

150

100

Fig. 28.11 (a) Control response for large look-ahead factor. (b) Con-

trol response for small look-ahead factor

The core activity is to observe and diagnose the

thermocouple readings to select the representative

temperature of the process as the process variable.

Cooperation among the thermocouples helps in de-

termining this information autonomously. It is very

important that the agents maintain close inspection of

the temperature sensors and ensure correct selection of

the representative temperatures to be able to set the

heating and cooling rates, as the material is very sus-

ceptible to these rates.

The control system has been implemented as a com-

bination of a proportional–integral–derivative (PID)

block with agents, as shown in Fig.28.10. The com-

bined operations of PID and agents constitute the main

control loop to generate heating and cooling rates as

the control response. The agents interact with the PID

in two main ways. First, to execute an accurate con-

trol response, the agents select the lead temperature

for the particular stage of the process, which could be

heating, cooling or neutral. The autoclave agent calcu-

lates the temperature set-point (T

set

) as an input to the

PID block. The temperature set-point is calculated to

be in the proximity of the upper or lower bound of the

proﬁle depending on the stage of curing. Second, the

thermocouple agents collect data from the curing model

(simulation) to select T

representative

. The PID block en-

sures that the process variable is driven closer to the

set-point value.

The PID block requires empirical calibration to

adjust the proportional, integral, and derivative coefﬁ-

cients of the equation. The tuning is performed online

while running the simulation and controller (for more

information on how totune PID coefﬁcients please refer

to the control system designliterature). The shape of the

curing proﬁle introduces too drastic inﬂections, making

it harder to maintain the steady-state condition. Nev-

ertheless, the introduction of the agents into the loop

allows for quicker and better utilization of the system

variables to cope with such changes.

The ability to look ahead into the future (greater

than 60s) was introduced into the agents to learn future

inﬂections in the curing proﬁle. The intention was to ad-

just the set-point to begin compensation early to prevent

overshoot. However, it was learned that the resultant

control response constrained the PID action by pro-

voking oscillation, as shown in Fig.28.11a. Throughout

experimentation, it was learned that the extent of the

look-ahead factor played a critical role in the instabil-

ity observations. The look-ahead factor was reduced

to a less than 30s horizon to make the PID ﬁltering

more dominant in the control of the overshooting, as

Part C 28.6

Distributed Agent Software for Automation 28.6 Composite Curing Results and Recommendations 483

showninFig.28.11b. Nonetheless, the inﬂexion points

are still a difﬁcult aspect to handle using the discussed

agent control algorithm. Without agent support, the

inﬂexion points that are introduced by the curing pro-

ﬁle will be conﬁned to a PID loop ﬁne-tuning task.

By expanding and contracting the look-ahead factor

of the agents, it is possible to perceive the effect of

the agent reasoning on the PID loop response. Ideally,

agents will act rapidly with a very short look-ahead

factor to provide better input to the PID loop. How-

ever, agents are not intended to carry out control;

they are designed to inﬂuence the control algorithm

in a beneﬁcial direction. An important lesson from

this experiment is the notion of role separation: that

which belongs to control must be in the control layer.

How to deﬁne this boundary is still an art rather than

a science.

Figure 28.11b shows an improved response, where

the temperature proﬁle follows the recipe with no oscil-

lations. According to speciﬁcations, the resultant cured

material from this experimental trial would be consid-

ered of high quality. To obtain an optimum solution

further adjustmentsto the PID coefﬁcients and agent ac-

tions are required and perhaps an expansion of the agent

TC02

(221)

2008-03-17 15:26:07

InformExecute(Success)

Actual CMs = 0

InformExecute(Success)

RequestRecruitAll

RequestPlan(ManageCuringOperations.on)

InformRecruitAll

RequestHandleEvent

InformHandleEvent

InformRecruitAll

InformExecute(Success)

Actual CMs = 0

RequestPlan($CuringEvents.startCuringProcess)

Actual CMs = 0

to TC08 from TC08

(222)

2008-03-17 15:26:07

(223)

2008-03-17 15:26:07

(224)

2008-03-17 15:26:07

(225)

2008-03-17 15:26:07

(226)

2008-03-17 15:26:07

(227)

2008-03-17 15:26:11

(228)

2008-03-17 15:26:11

(229)

2008-03-17 15:26:11

(230)

2008-03-17 15:26:12

(231)

2008-03-17 15:26:12

(232)

2008-03-17 15:26:12

(233)

2008-03-17 15:26:12

ACSObj1 Autoclave01 Curing GlobalDF1 TC01

from TC08

to TC12

RequestPlan(ManageCuringOperations.on)

to TC05

RequestPlan(ManageCuringOperations.on)

to TC06

RequestPlan(ManageCuringOperations.on)

to TC16

RequestPlan(ManageCuringOperations.on)

to TC09

$CuringEvents.start...

Curing

$CuringEvents.start...

ACSObj1

ManegeCuringOpera...

TC12

ManegeCuringOpera...

TC09

ManegeCuringOpera...

TC16

ManegeCuringOpera...

TC06

ManegeCuringOpera...

TC02

ManegeCuringOpera...

TC05

ManegeCuringOpera...

TC04

ManegeCuringOpera...

TC13

ManegeCuringOpera...

TC11

ManegeCuringOpera...

TC15

ManegeCuringOpera...

Autoclave01

ManegeCuringOpera...

TC01

ManegeCuringOpera...

TC07

ManegeCuringOpera...

TC14

ManegeCuringOpera...

TC03

ManegeCuringOpera...

TC10

ManegeCuringOpera...

TC08

Details

RequestHandleEvent

InformHandleEvent

Fig. 28.12 Agent communication

intelligence. Nevertheless, these improvements fall out-

side the scope of this chapter.

During the agent cooperation, the agents talk to

each other using the FIPA/JDL messaging protocol.

Figure 28.12 shows an example agent communication.

The observer can trace the agent transactions to learn

what decisions have been made. The transactions show

the information that the agents exchanged to arrive at

a control decision. In Fig.28.12 a multicasting exam-

ple is shown, where a single agent contacts multiple

agents in a two-way transaction. In multiagent systems,

interagent communication consists of request/inform

transactions, where an agent transmits a request for in-

formation to one or a group of agents. The receiving

agents process the request locally but may initiate con-

sultation with other agents as well by using similar

request/inform transactions. Once the receiving agent

has ﬁnished preparing its response, it uses an inform

message to send the information back to the requester.

The requester then receives multiple responses from the

agents. Each response provides a different perspective

of the situation under consultation. The requester must

select the most valuable information or a combination

of it to conclude its own decision-making cycle.

Part C 28.6

484 Part C Automation Design: Theory, Elements, and Methods

The distributed nature of agent messaging allows

for a realistic expansion of the system to add/remove

functionality as desired.

In the current implementation, all agents reside

within one controller, but the agents can be distributed

in any desired topology. The distribution topology is

a design decision that must comply with the applica-

tion’s characteristics and requirements. A rule of thumb

to decide on the agent distribution is to weigh the

desired degree of survivability, diagnosability, and re-

conﬁgurability of the system.

In a distributed agent implementation, agents reside

in separate platforms. The distribution of the agents

depends on the partitioning criteria selected for the

application. The partitioning criteria takes under con-

siderations the number of messages used in negotiation.

The idea is to use a reduced number of messages.

A poor division of functionality among the agents will

result in more messaging among them. Message traf-

ﬁc and bandwidth utilization are good metrics to deﬁne

optimum partitioning.

28.6.3 Timing and Stability Criteria

The multiagent system guarantees an exhaust search for

an optimum conﬁguration within a speciﬁc time hori-

zon. For small systems(< 100agents), it isfairlysimply

to establish the time horizon required by the agents to

accomplish a complete search of conﬁgurations. The

time horizon estimation is established by injecting var-

ious conditions into the simulation model so the system

designer can observe what time is required to search

for and converge to a solution. Once this time hori-

zon is estimated, each agent is given a factored time

limit to participate in a particular negotiation process.

An inﬁnite time horizon is never given to the agents

to avoid deadlocks. Unknown singularities could gen-

erate deadlocks in the system. It is guaranteed that the

agents will ﬁnd a solution within the predeﬁned time

horizon. However, the value of the solution will ﬂuc-

tuate within optimum and-near optimum performances,

depending on the complexity of the emergent search

space.

28.7 Conclusions

This chapter has shown how to use industrial agent

technology to solve a composite curing application.

It has also shown some of the technology require-

ments and implications of extending industrial PLCs

with agent ﬁrmware. An example application was

used to show the design and modeling phases and

the validation of the solution using simulation. Im-

portant results and recommendations can be extracted

from this to help the modernization of curing tech-

nology. Using interagent messaging and the ability to

cooperate on the creation and assessment of thermo-

couple conﬁgurations, the agents enable a dynamic

learning system, always adapting to current condi-

tions.

The PLC ﬁrmware provides for a multitasking

priority-based executive control cycle. Agents on built

as user level tasks to execute at a low priority. Funda-

mental tasks in the PLC that manage the PLC’s integrity

are never interrupted by the agent tasks. User-deﬁned

tasks such as stand-alone periodic and continuous con-

trol loops compete with agent tasks, but based on their

priority levels. Among these, periodic tasks are higher

order than agents, but never higher than the executive

functions.

The extent of learning depends on the sophistica-

tion of the rules given to the agents. This aspect is

a user-deﬁned characteristic of the system. Agent pro-

gramming provides for this type of ﬂexibility. It has

been learned that loosely coupled agents and control

components ease the scalability of the control solution.

Finally, this chapter has shown how survivability, diag-

nostics, and reconﬁguration of a control system can be

achieved with component-level intelligence in a more

compact manner.

28.8 Further Reading

1. F. Bergenti, M.P. Gleizes, F. Zambonelli (Eds.):

Methodologies and Software Engineering for Agent

Systems: The Agent-Oriented Software Engineering

Handbook (Springer, New York 2004)

Part C 28.8