![](bg3.png)
estimation neural network on line; SCAs are used to
cut off failed sensors.
Figure 2: Architecture of MAS based FTC system.
In this FTC system there are three distinct parts that
differ from traditional control systems: fault
diagnosis agents to diagnose system failure, state
estimation agents to estimate state approximately
when the sensor failed, and system self-adaptive
controller, realized by FTCA, for FTC under
actuator faults:
(1) In the LIF based FDA, two kinds of faults,
actuator faults and sensor faults will be diagnosed. If
actuator faults happened, FTCA will take some
responding actions according to the fault state.
While the sensor faults have been found, the SEA
will play a part in the sensors.
(2) A SEA based on an output recurrent neural
network is proposed to provide the system states
estimation to replace the fault sensors. Usually, SCA
controls the switch to be on the position 2. In this
case, the system state feedback signals will be from
sensors. If the sensor faults occur, SCA will switch
to the position 1. In this case, the signals from SEAs
will be used as the system feedback.
(3) The FTCA, system controller, will give
different control strategy according to the system
state and fault cases to keep the system performance.
For example, when a fault from the actuator was
detected, the FTCA based on AFNN could adjust the
control signal to overcome the influence of the fault
according to the system response on line.
3 INTELLIGENT ALGORITHMS
OF AGENTS
As mentioned above, algorithms encapsulated in an
agent’s IPM determines its functions. In the IPM of
a FDA, multi-layer information fusion technology is
adopted for fault diagnosis, which separated fault
diagnosis into two parts: local diagnosis fusion
implemented by multi-sensors fuzzy inference and
global diagnosis fusion implemented by a three-
layer fuzzy neural network. In SEA, a new output
recurrent neural network is designed to construct the
system state estimator. For FTCA, a self-adaptive
fuzzy neural network is proposed as its information
process method. Detail about these solutions could
refer to (Yao, 2006)
4 SIMULATION EXPERIMENTS
4.1 Experiment Platform
In this simulation, the MAS framework is coded in
JADE platform, a software development framework
for agent application developed by TILAB. The
algorithms mentioned in section 3 are coded in
Matlab 6.5.
To implement calling Matlab methods from
JADE, JMatLink, a small toolkit to connect Java
with Matlab, is used to call for the functions in an
m-file.
The main user interface of MAS compiled in
JADE is shown in figure 3.
4.2 Working Flow of MAS
A prominent advantage of MAS is that agents could
discover an optimized way to fulfil tasks by
negotiation, coordination and cooperation via
sending messages. Accordingly, the communication
and cooperation between agents are the most two
important research topics regarding MAS.
The communication among agents, as well as
their knowledge and mental state, is based upon
domain ontology. But different ontologies regarding
one domain may exist sometimes in a system. These
ontologies contain different terms, which engender
great obstacles for agent communication, to express
same or similar concepts. Dealing with this problem,
a method called term substitution based on
intelligent ontology mapping is proposed and
implemented by ontology mediation agent (OMA).
OMA maintain glossary mapping tables between
domain ontologies. When an agent receives an ACL
message containing a few baffling words, the agent
forwards this message to OMA for interpretation. If
OMA could find terms in an ontology upon which
this agent based corresponding to those baffling ones,
it will substitute them and send the message back.
And then, the agent will understand the message
meaning.
AFNN-Adaptive Fuzzy Neural Network
ORNN-Output Recurrent Neural Network
LIF- Layered Information Fusion
A MARINE FAULTS TOLERANT CONTROL SYSTEM BASED ON INTELLIGENT MULTI-AGENTS
359