Rejection of sensor deterioration, noise, disturbance and plant parameters
1 like141 views
1) The document describes a proposed fault tolerant controller system (FTCS) strategy for temperature control of an air conditioning system. The FTCS is designed to be robust against noise, sensor failures, disturbances, and parameter variations while maintaining optimal performance. 2) A mathematical model is developed consisting of models for the heat exchanger and conditioning space. A neural network model is also developed to model the system dynamics. 3) The neural network is trained using a backpropagation algorithm to minimize the mean squared error between the network outputs and actual system outputs over time. Training results show the network can accurately model the system.
![The Sec
Malacca
out
1,2
f
1. Intr
Since th
isolation (F
faults in pr
problems b
and isolati
actions to
indicators
is present,
not, by itse
performanc
to eliminat
controller (
2. Math
There a
mathemati
laws of n
process [
mathemati
first part
system) an
room. Thes
2.1 Hea
Based
thermal ba
The tran
be derived
cond Internatio
a, Malaysia, Ju
Outside ro
Heat exch
Indoor and
roduction
he beginning o
FDI) has been
rocess plants [
but it seems c
ing the fault, t
o solve the p
may reach alar
even if it is d
elf, keeps the s
ce level, but h
te the fault and
(FTC) is solvin
hematical M
are basically
cal model of
nature or thro
[2]. The fir
cal model for H
is the air-pro
nd the second
se two parts ar
Figure (1) Room
at Exchanger m
on the law
alance equation
nsfer function
from equ. (1)
onal Conferenc
une 2-3, 2009
oom
hanger, room re
d outdoor
of the 70's, fau
n used as a to
[1].FDI can so
clear that apart
the FDI schem
problem. Qua
rming levels ev
detected by the
system working
human interven
d its effects. T
ng this problem
Model for HV
two ways of
f a system: by
ough experime
rst one will
HVAC which c
ocessing unit
d part is the
re illustrated in
m-air conditionin
model
conservation
n is given as fol
.......................
of handled air
as fallow.
ce on Control, I
egion
ult detection an
ol on managin
olve some of th
t from detectin
me does not tak
ality and safe
ven when a fau
e FDI. FDI do
g at the require
ntion is require
The fault tolera
m.
VAC System
f determining
y implementin
entation on th
l develop th
contain from th
(heating/coolin
air-conditionin
n figure (1).
ng system
of energy, th
llowing [4]:
..............(1)
temperature ca
Instrumentation
nd
ng
he
ng
ke
ety
ult
es
ed
ed
ant
m
a
ng
he
he
he
ng
ng
he
an
Wh
2.2
H
model
the sp
exchan
temper
heat ca
the spa
similar
Her
handlin
assume
conditi
supply
conditi
Wh
The
conditi
3. N
The
them r
n and Mechatr
here
,
Conditioning
Here the suppo
of air-conditio
ace is airtight
nge between In
rature in the s
apacity of the
ace is ignored
r to the heat ex
re the supply
ng air tempera
e that the spe
ioning room is
y air and Cr
ioning space te
here
,
e transfer fu
ioning space ca
Neural Netw
e neural netwo
remarkable an
ronic Engineeri
.....
g space model
osition for the
oning space is
t and there is
ndoors and ou
space is almos
door, window
d. The thermal
xchanger [2].
air temperatu
ature after AH
ecific heat cap
s equal to spec
= Ca. The t
emperature can
.......
,
unction of he
an be got from
work Model
rks have many
nd none more
ing (CIM09)
.....................(2
,
temperature c
s followed as:
not the direc
utdoors; second
st equal; third
ws and the go
balance equat
ure to room
HU and Ts= Tl
pacity of air i
cific heat capac
transfer functi
n be got from (
.......................
eat exchanger
m (2) and (4)
.........
l
y properties, so
important tha
2)
control
firstly
ct heat
dly the
dly the
ods in
tion is
(3)
is the
. Here
in air-
city of
ion of
3).
(4)
r and
(5)
ome of
an the
605](https://image.slidesharecdn.com/rejectionofsensordeteriorationnoisedisturbanceandplantparameters-170910212549/85/Rejection-of-sensor-deterioration-noise-disturbance-and-plant-parameters-2-320.jpg)
![The Sec
Malacca
“universal
neural netw
structures
approxima
tractable s
problems[5
dynamics
networks.
exogenous
network, w
layers of t
the linear
time-series
NARX mo
, … ,
Where u
network at
order, and
the functio
Perceptron
recurrent n
network stru
Figure
Neural n
need to be
numerical
desired beh
training al
closely to
consists of
samples, w
has 4000
(validation
several pas
is shown in
cond Internatio
a, Malaysia, Ju
function appr
works usually
to exploit
ation property
olutions to the
5]. All the dy
only at the
The nonlinear
s inputs (NAR
with feedback
the network. T
ARX model,
s modeling. T
odel is [6].
,
1 .........
u(t) and y(t) re
t time t, Du an
the function f
on f can be a
n, the resultin
neural networ
ucture with four
e (2) A NARX n
nets contain co
e taught (train
(data set) e
havior) while
lgorithm until
the examples
f two batches;
which are used
samples, w
n). To include p
st output value
n equation (6).
onal Conferenc
une 2-3, 2009
roximation pro
used with no
their “univ
y” in this wa
e posed nonlin
ynamic netwo
input layer,
r autoregressiv
RX) is a rec
connections en
The NARX mo
which is com
The defining e
… , 1 ,
.......................
epresent input a
nd Dy are the i
f is a nonlinear
approximated b
ng system is
rk. Figure (2)
output delays.
network with fou
oefficients calle
ning) by being
examples (that
the weights ar
l the neural n
as possible [7
; one batch ha
d for training;
which are us
plant dynamics
es and past inpu
ce on Control, I
operty” therefo
onlinear netwo
versal functio
ay to compu
near for HVA
orks have eith
or feedforwa
ve network wi
current dynam
nclosing sever
odel is based o
mmonly used
equation for th
................(6)
and output of th
input and outp
r function. Whe
by a Multilay
called a NAR
show the NAR
ur output delays.
ed 'weights' the
g presented wi
t represent th
re modified by
net performs
7]. The data s
as 4000 rando
the other batc
sed for testin
s, the model us
ut values, whic
Instrumentation
ore
rk
on
ute
AC
her
ard
ith
mic
ral
on
in
he
he
put
en
yer
RX
RX
ey
ith
he
y a
as
set
om
ch
ng
es
ch
3.1
Tra
most t
develo
trainin
proble
includi
data p
and bia
Neu
weight
networ
given
algorit
networ
Dyn
self-su
has an
differe
differe
iteratio
compu
by the
criterio
chosen
SSE fo
for a w
To
regress
The
algorit
where
Fig
n and Mechatr
Neural netwo
aining the artif
time-consumin
opment. It is v
ng algorithm
m, as succes
ing the comple
oints in the tr
ases in the netw
ural network tr
ts (w) and th
rk until the b
performance o
thm is back-p
rks). But in thi
namic neural n
ustained activit
n activation lev
ent neurons c
ent changes o
on the Sum
uted to check w
e update weig
on value α or
n to be a smal
or the new net
well-trained net
∑
compute m
sion: MSE = s
e result of th
thm through i
α value chosen
gure (3) Mean Sq
ronic Engineeri
ork model trai
ficial neural ne
ng task in th
very difficult
will be the f
s will depend
exity of the pro
aining set, the
work, and the e
raining is the a
he bias terms
behavior obtai
objective. The
propagation (S
s model used d
networks are c
ty patterns, in
vel that can ch
can be at dif
over time. Du
of Squared
whether the ne
ght matrices i
r not. Here, th
ll number that
twork weight.
twork is that [8
mean-squared
sum (SSE)/leng
he NARX pl
iteration is sh
n as 1×10-10
.
quare Error thro
ing (CIM09)
ining
etwork (ANN)
he ANN appli
to recognize
fastest for a
d on many fa
oblem, the num
e number of w
error goal [10]
activity of fixin
(b) throughou
ined achieves
e most used tr
Static (feedfor
dynamic netwo
capable of prol
which each n
hange over tim
fferent levels,
uring each tr
Errors (SSEs
twork output c
is smaller tha
he criterion va
t is acceptable
Thus, the con
8]:
.....................
error (MSE
gth (iteration)
lant model tr
hown in figur
ough 1000 iterati
is the
ication
which
given
actors,
mber of
weights
.
ng the
ut the
some
raining
rward)
ork;
longed
neuron
me, and
, with
raining
s) are
caused
an the
alue is
as an
ndition
(7)
E) for
raining
re (3)
on
606](https://image.slidesharecdn.com/rejectionofsensordeteriorationnoisedisturbanceandplantparameters-170910212549/85/Rejection-of-sensor-deterioration-noise-disturbance-and-plant-parameters-3-320.jpg)
![The Sec
Malacca
The tra
input sign
shown in
However, t
Therefore,
the obtaine
Figure (4)
3.2 Neu
Afterwa
model val
estimated
use of the
data well o
The inp
plant and t
(5). It clear
output.
Figure (5)
cond Internatio
a, Malaysia, Ju
aining carried
al to the mod
figure (4), th
this could be c
the validation
ed model.
) training output
ural network m
ards a model
lidation. This
model is suffi
model (to che
or not) [9].
put-output data
the model are c
rly shows that
validation outpu
onal Conferenc
une 2-3, 2009
out by applyin
del and check
he model fit t
caused by over-
n tests are nex
t Error between p
model validati
identified, th
step is to te
iciently good f
ck if the mode
a set obtained
compared and
the model outp
ut Error between
ce on Control, I
ng 4000 samp
the output. A
the training se
-fitting problem
xt performed o
plant and model
ion
he next step
est whether th
for the intende
el can fit the te
d from both th
shown in figu
put fit the actu
n plant and mode
Instrumentation
ple
As
et.
m.
on
is
he
ed
est
he
ure
ual
el
4. Ar
The
system
interco
Figu
Wh
conditi
networ
represe
gain r
contro
feedba
the sy
the re
sensor
5. S
To eva
implem
presen
The pa
conditi
height,
kJ/kg.C
criterio
field, t
room a
Ga=1.0
1/R=0
cold-w
summe
middle
follow
parame
n and Mechatr
rchitecture o
e proposed mo
m setup is
onnection in fig
ure (6) Block dia
here the heat ex
ioning space
rk model) re
ents the tradit
represent the v
l system imple
ack with PID c
stem stable an
esponse speed,
r deterioration a
Simulation
aluate the good
mented on the H
nted below.
arameters of th
ioning room is
, the air specifi
Co
, the air dens
on in the heatin
the number of t
and the calcula
08kg/s, the hea
.2kw/Co
and th
water and back-
er is Twin-Twout=
e process, heat
wing: K1= -19.3
eters of air-con
ronic Engineeri
of the FTC
odel based on f
shown as
gure (6).
agram of propose
structure
xchanger repre
represents G
epresent the
tional controll
variant gain f
ements the com
control. The PI
nd reject the d
, reject the n
and reduce the
dness of the pro
HVAC system
he HVAC syste
s 10 m length,
ic heat capacity
sity is 1.2 kg/m
ng, ventilation
taking a breath
ation of the sup
at resistance of
he temperature
-water to the he
= - 5Co
. At last
exchanger, are
35Co
. s /kg, T1=
nditioning room
ing (CIM09)
fault tolerant c
a block di
ed model based F
esents G1(s), th
G2(s), NNM (
plant model,
ler and the s
for the sensor
mbination of d
ID control can
disturbance, qu
noise, overcom
overshoot.
oposed techniq
m, the simulatio
em: the volume
8m width and
y is Ca=1.0
m3
, Based on th
and air-condit
h in air-conditio
pply air is
f wall is
error of supply
eat exchanger
t the parameter
e calculated as
=30s . The
m are calculate
control
iagram
FTC
he air-
(neural
, PID
ensors
r. The
double
n make
uicken
me on
ques,
n is
e of air
4.5m
he
ioning
oning
y
in
rs of
ed as
607](https://image.slidesharecdn.com/rejectionofsensordeteriorationnoisedisturbanceandplantparameters-170910212549/85/Rejection-of-sensor-deterioration-noise-disturbance-and-plant-parameters-4-320.jpg)
![The Sec
Malacca
following:
object G(s)
and the
be express
Simula
FTC and
validate th
control in t
K(s) to 0.4
result. It is
don’t effe
there’s ef
controller
the simul
temperatur
Figure 7
Simula
FTC and
certify the
the contro
simulate w
gain of
K1K2=9.28
T2=406s, w
object G(s)
G s
cond Internatio
a, Malaysia, Ju
K2=0.4Co s /k
) can express a
.
,
.
disturbances
ed as:
0.052
0.81 0.
tion (1) Senso
PID Controlle
he rejection to
the HVAC sys
4 i.e. sensor fau
s seen that the
ected with d
ffectiveness o
(Note. The res
lation have
re at 20 Co
i.e.
7 Sensors deterio
tion (2) Param
PID Controlle
robust of FT
olled process
while the contro
controlled pr
8 and the time
which both in
) can express a
.
onal Conferenc
une 2-3, 2009
kg, T2=338s. So
as: [2]
.
.................
from outdoor
926
or Deterioratio
er with norma
o the sensor de
tem we change
ult = 0.4 figur
e output of FT
deterioration s
on the system
sults of the res
been normal
4 mV indicate
oration effectiven
meter Variatio
er with norma
control in the
parameters a
ollers are kept
rocess, K1K2=
e constant, T2=
ncrease 20%.
as:
,
ce on Control, I
o the controlled
.........(8)
..............(9)
and indoors ca
on Rejection f
al behavior. T
eteriorate of F
e the gain sens
e 7 illustrate th
T control syste
sensor where
m in the PI
sponse shown
lized to roo
s 20 Co
).
ness on the syste
on Rejection f
al behavior. T
HVAC system
are changed
unchanged. Th
=7.74, becom
=338s, becom
The controlle
.
Instrumentation
d
an
for
To
FT
or
he
em
as
ID
in
om
em
for
To
m,
to
he
mes
mes
ed
The
Fi
Sim
Contro
that t
effecti
system
Fig
Sim
PID C
rejectio
HVAC
proces
contro
(1). Fi
nomin
Fig. 10
6-
In
been p
and pla
n and Mechatr
e simulation re
igure 8 paramete
mulation (3) N
oller with norm
the controller
veness while P
m.
gure 9 the noise e
mulation (4) D
Controller with
on to the dis
C system, the
ss increases 0.1
llers paramete
ig. 10 show FT
al PID and rob
0 comparison of
Conclusion
this paper, a
proposed based
ant parameter
ronic Engineeri
sults are shown
er variation and t
Noise Rejectio
mal behavior.
r FTC remai
PID lose posse
effectiveness on
Disturbance Rej
normal behav
sturbances of
input disturb
1 in ramp at 0
rs are kept as
TC with distur
bust.
disturbance effe
nominal PID
fault-tolerant
d on rejection
variation. This
ing (CIM09)
n in figure 8.
the robust of FT
n for FTC an
Figure (9) is s
ins robust w
ess of the cont
n the PID control
ejection for FT
vior. To valida
FT control i
bance of cont
0th second whi
same as Simu
rbance is faste
ects on the system
control schem
sensor deterio
s controller des
TC
d PID
shown
without
trol on
ller
TC and
ate the
in the
trolled
ile the
ulation
er than
m with
me has
oration
sign is
608](https://image.slidesharecdn.com/rejectionofsensordeteriorationnoisedisturbanceandplantparameters-170910212549/85/Rejection-of-sensor-deterioration-noise-disturbance-and-plant-parameters-5-320.jpg)
![The Second International Conference on Control, Instrumentation and Mechatronic Engineering (CIM09)
Malacca, Malaysia, June 2-3, 2009
compatible with HVAC systems where the traditional
PID controller unable to deal with plant has nonlinear,
pure lag and high inertia that are presented in the
HVAC system. Thereby FTC more efficient than PID;
Increased efficiency tends to reduce the cost associated
with operating the HVAC system.
7. References
[1] Silva, P.M.; Becerra, V.M.; Khoo, I.; Calado, J.M.F
(2006) “Multiple-Model Fault Tolerant Control of Terminal
Units of HVAC Systems” International Symposium on
Industrial Electronics, Volume 4, 9 July 2006 Page(s):2896 -
2901 (2006)
[2] Wang, Jiangjiang; Zhang, Chunfa; Jing, Youyin;
(2007)”Hybrid CMAC-PID Controller in Heating Ventilating
and Air-Conditioning System” International Conference on
Mechatronics and Automation, Pp.3706 – 3711, 5-8 Aug.
2007
[3] Omid Omidvar ; david L. Elliott (1997)”Neural
systems for Control” Chp. 5 Pp.93-101 February (1997)
[4] C. Li, W. Zhang and Y. Liang, “The Decoupling
control of constant temperature and humidity air-
conditioning system,” Refrigeration and Air-conditioning”,
vol. 6, no. 4, pp 42-47, August 2006.
[5] Zdzislaw Bubnicki (2005) “Modern Control Theory”
[6] (L.R. Medsker and L.C. Jain 2001) ” Recurrent Neural
Networks”
[7] JR Leigh (2004) “Control Theory” Second Edition
chp.17 Pp.225-228 (2004)
[8] Haider Abbas Fadhl Mohamed (1998) thesis
“Application of Neural Network for Modeling and Control in
Induction Machine”
[9] Y. Zhu (2001)” Multivariable System Identification
for Process Control” October 2001, Publisher: Elsevier
Science & Technology Books
[10] Paul S,erban Agachi, Zolt_n K. Nagy, Mircea Vasile
Cristea, and Arpad Imre-Lucaci (2006) “Model Based
Control” Chp. 1 Pp.14 publisher WILEY-VCH Verlag
GmbH & Co. KGaA
8. BIOGRAPHIES
Mr. Raad Z. Homod He
graduated with a B.Eng. in
Mechanical Engineering from the
University of Basrah, Iraq in
1991. He worked as engineer for
five years in General
Establishment of Steel & Iron
(Iraq) and then he worked ten years as HVAC
engineering in Al-Tomoh Al-Kabir Company (Libya).
Currently, he is perusing his Master in HVAC Systems
and Control at the Department of Mechanical
Engineering in University of Malaya.
DR. T.M. Indra Mahlia
received B.Eng from the
University of Syiah Kuala,
Indonesia, and M.Eng.Sc and
PhD from the University of
Malaya. He is currently an
Associate Professor at the
Department of Mechanical
Engineering, University of Malaya, Kuala Lumpur,
Malaysia
Dr. Haider A. F. Mohamed
received his PhD in Electrical
Engineering from the
University of Malaya,
Malaysia in 2006. He worked
as a computer engineer for
two years and as a researcher
for four years before he
became a lecturer in the
Department of Electrical
Engineering in University of Malaya, Malaysia, in
2000. His main research fields are identification and
nonlinear intelligent control of various systems such as
robot arms, automated guided vehicles, and electric
drives.
609](https://image.slidesharecdn.com/rejectionofsensordeteriorationnoisedisturbanceandplantparameters-170910212549/85/Rejection-of-sensor-deterioration-noise-disturbance-and-plant-parameters-6-320.jpg)
Rejection of sensor deterioration, noise, disturbance and plant parameters
- 1. The Second International Conference on Control, Instrumentation and Mechatronic Engineering (CIM09) Malacca, Malaysia, June 2-3, 2009 Rejection of Sensor Deterioration, Noise, Disturbance and Plant Parameters Variation in HVAC System Raad Z. Homod, T. M. I. Mahlia, 1 Haider A. F. Mohamed Center for research in Applied Electronics (CRAE) University of Malaya, 50603 Kuala Lumpur, Malaysia Tel: +60133353502 Email: raadahmood@yahoo.com 1 Department of Electrical & Electronic Engineering The University of Nottingham Malaysia Campus Jalan Broga, 43500 Semenyih, Selangor Darul Ehsan, Malaysia E-mail: haider.abbas@nottingham.edu.my Abstract The main objective of this paper is to reject faults tolerant like sensor deterioration, noise, disturbance and depreciation of the plant in Heating, Ventilating and Air Conditioning (HVAC) System. a fault tolerant controller system (FTCS) strategy for temperature control of air-conditioning system has been proposed. The setpoint response controller in the FTCS is designed in terms of insusceptible with noise, failure sensor, disturbance and parameter variation with optimal performance specification. According to the system operation requirement for faults rejection, a closed-loop for rejecting fault signals is configured. It is found that the setpoint response, the disturbance rejection, the noise rejection, the robust, rejection of failure sensor and the performance of the designed FTCS are better than conventional PID. Therefore the residual signal can now be used to monitor faults in the system while the proposed method ensures correctness of operation. Keywords- Noise rejection, HVAC system, disturbance rejection, plant parameters variation, fault tolerant controller, Sensor deterioration. NOMENCLATURE Symbols Mh The quality of heat exchanger, (kg) AHU Air handling unit Mr Air quality of air-conditioning room, (kg) Ga Supply air flows ,(kg/s) Gw Cold water (or hot water) flows,(kg/s) Ca Specific heat capacities of air, (kj/kg Co ) Cw Specific heat capacities of water(kj/kg Co ) Ch Specific heat capacities of heat exchanger (kj/kg Co ) Tm The temperature of mix fresh air, (Co ) Tl The temperature after heat exchanger, (Co ) Th Surface temperature of heat exchanger,(Co ) Twin The temperature of supply water, (Co ) Tw out Back-water from heat exchanger, (Co ) Qroom Perturbations inside thermal load, (kj) Tout Uncontrolled outside temperature (Co ) K1 the amplify coefficient of heat exchanger, (Co .s /kg) T1 Heat exchanger time constant, (s) Gr(s) the disturbance of heat exchanger, (kg/s) K2 the amplify coefficient of room, (Co .s/kg); is; is Here, T2 Conditioning space time constant, (s) Tf (s) the disturbance to room, which include the disturbances from outdoor and indoor, (Co ). Subscripts h Heat exchanger r Room a Air w Water m Mixed fresh and return air l Leaving W in Water input W out Water output room Inside room 604
- 2. The Sec Malacca out 1,2 f 1. Intr Since th isolation (F faults in pr problems b and isolati actions to indicators is present, not, by itse performanc to eliminat controller ( 2. Math There a mathemati laws of n process [ mathemati first part system) an room. Thes 2.1 Hea Based thermal ba The tran be derived cond Internatio a, Malaysia, Ju Outside ro Heat exch Indoor and roduction he beginning o FDI) has been rocess plants [ but it seems c ing the fault, t o solve the p may reach alar even if it is d elf, keeps the s ce level, but h te the fault and (FTC) is solvin hematical M are basically cal model of nature or thro [2]. The fir cal model for H is the air-pro nd the second se two parts ar Figure (1) Room at Exchanger m on the law alance equation nsfer function from equ. (1) onal Conferenc une 2-3, 2009 oom hanger, room re d outdoor of the 70's, fau n used as a to [1].FDI can so clear that apart the FDI schem problem. Qua rming levels ev detected by the system working human interven d its effects. T ng this problem Model for HV two ways of f a system: by ough experime rst one will HVAC which c ocessing unit d part is the re illustrated in m-air conditionin model conservation n is given as fol ....................... of handled air as fallow. ce on Control, I egion ult detection an ol on managin olve some of th t from detectin me does not tak ality and safe ven when a fau e FDI. FDI do g at the require ntion is require The fault tolera m. VAC System f determining y implementin entation on th l develop th contain from th (heating/coolin air-conditionin n figure (1). ng system of energy, th llowing [4]: ..............(1) temperature ca Instrumentation nd ng he ng ke ety ult es ed ed ant m a ng he he he ng ng he an Wh 2.2 H model the sp exchan temper heat ca the spa similar Her handlin assume conditi supply conditi Wh The conditi 3. N The them r n and Mechatr here , Conditioning Here the suppo of air-conditio ace is airtight nge between In rature in the s apacity of the ace is ignored r to the heat ex re the supply ng air tempera e that the spe ioning room is y air and Cr ioning space te here , e transfer fu ioning space ca Neural Netw e neural netwo remarkable an ronic Engineeri ..... g space model osition for the oning space is t and there is ndoors and ou space is almos door, window d. The thermal xchanger [2]. air temperatu ature after AH ecific heat cap s equal to spec = Ca. The t emperature can ....... , unction of he an be got from work Model rks have many nd none more ing (CIM09) .....................(2 , temperature c s followed as: not the direc utdoors; second st equal; third ws and the go balance equat ure to room HU and Ts= Tl pacity of air i cific heat capac transfer functi n be got from ( ....................... eat exchanger m (2) and (4) ......... l y properties, so important tha 2) control firstly ct heat dly the dly the ods in tion is (3) is the . Here in air- city of ion of 3). (4) r and (5) ome of an the 605
- 3. The Sec Malacca “universal neural netw structures approxima tractable s problems[5 dynamics networks. exogenous network, w layers of t the linear time-series NARX mo , … , Where u network at order, and the functio Perceptron recurrent n network stru Figure Neural n need to be numerical desired beh training al closely to consists of samples, w has 4000 (validation several pas is shown in cond Internatio a, Malaysia, Ju function appr works usually to exploit ation property olutions to the 5]. All the dy only at the The nonlinear s inputs (NAR with feedback the network. T ARX model, s modeling. T odel is [6]. , 1 ......... u(t) and y(t) re t time t, Du an the function f on f can be a n, the resultin neural networ ucture with four e (2) A NARX n nets contain co e taught (train (data set) e havior) while lgorithm until the examples f two batches; which are used samples, w n). To include p st output value n equation (6). onal Conferenc une 2-3, 2009 roximation pro used with no their “univ y” in this wa e posed nonlin ynamic netwo input layer, r autoregressiv RX) is a rec connections en The NARX mo which is com The defining e … , 1 , ....................... epresent input a nd Dy are the i f is a nonlinear approximated b ng system is rk. Figure (2) output delays. network with fou oefficients calle ning) by being examples (that the weights ar l the neural n as possible [7 ; one batch ha d for training; which are us plant dynamics es and past inpu ce on Control, I operty” therefo onlinear netwo versal functio ay to compu near for HVA orks have eith or feedforwa ve network wi current dynam nclosing sever odel is based o mmonly used equation for th ................(6) and output of th input and outp r function. Whe by a Multilay called a NAR show the NAR ur output delays. ed 'weights' the g presented wi t represent th re modified by net performs 7]. The data s as 4000 rando the other batc sed for testin s, the model us ut values, whic Instrumentation ore rk on ute AC her ard ith mic ral on in he he put en yer RX RX ey ith he y a as set om ch ng es ch 3.1 Tra most t develo trainin proble includi data p and bia Neu weight networ given algorit networ Dyn self-su has an differe differe iteratio compu by the criterio chosen SSE fo for a w To regress The algorit where Fig n and Mechatr Neural netwo aining the artif time-consumin opment. It is v ng algorithm m, as succes ing the comple oints in the tr ases in the netw ural network tr ts (w) and th rk until the b performance o thm is back-p rks). But in thi namic neural n ustained activit n activation lev ent neurons c ent changes o on the Sum uted to check w e update weig on value α or n to be a smal or the new net well-trained net ∑ compute m sion: MSE = s e result of th thm through i α value chosen gure (3) Mean Sq ronic Engineeri ork model trai ficial neural ne ng task in th very difficult will be the f s will depend exity of the pro aining set, the work, and the e raining is the a he bias terms behavior obtai objective. The propagation (S s model used d networks are c ty patterns, in vel that can ch can be at dif over time. Du of Squared whether the ne ght matrices i r not. Here, th ll number that twork weight. twork is that [8 mean-squared sum (SSE)/leng he NARX pl iteration is sh n as 1×10-10 . quare Error thro ing (CIM09) ining etwork (ANN) he ANN appli to recognize fastest for a d on many fa oblem, the num e number of w error goal [10] activity of fixin (b) throughou ined achieves e most used tr Static (feedfor dynamic netwo capable of prol which each n hange over tim fferent levels, uring each tr Errors (SSEs twork output c is smaller tha he criterion va t is acceptable Thus, the con 8]: ..................... error (MSE gth (iteration) lant model tr hown in figur ough 1000 iterati is the ication which given actors, mber of weights . ng the ut the some raining rward) ork; longed neuron me, and , with raining s) are caused an the alue is as an ndition (7) E) for raining re (3) on 606
- 4. The Sec Malacca The tra input sign shown in However, t Therefore, the obtaine Figure (4) 3.2 Neu Afterwa model val estimated use of the data well o The inp plant and t (5). It clear output. Figure (5) cond Internatio a, Malaysia, Ju aining carried al to the mod figure (4), th this could be c the validation ed model. ) training output ural network m ards a model lidation. This model is suffi model (to che or not) [9]. put-output data the model are c rly shows that validation outpu onal Conferenc une 2-3, 2009 out by applyin del and check he model fit t caused by over- n tests are nex t Error between p model validati identified, th step is to te iciently good f ck if the mode a set obtained compared and the model outp ut Error between ce on Control, I ng 4000 samp the output. A the training se -fitting problem xt performed o plant and model ion he next step est whether th for the intende el can fit the te d from both th shown in figu put fit the actu n plant and mode Instrumentation ple As et. m. on is he ed est he ure ual el 4. Ar The system interco Figu Wh conditi networ represe gain r contro feedba the sy the re sensor 5. S To eva implem presen The pa conditi height, kJ/kg.C criterio field, t room a Ga=1.0 1/R=0 cold-w summe middle follow parame n and Mechatr rchitecture o e proposed mo m setup is onnection in fig ure (6) Block dia here the heat ex ioning space rk model) re ents the tradit represent the v l system imple ack with PID c stem stable an esponse speed, r deterioration a Simulation aluate the good mented on the H nted below. arameters of th ioning room is , the air specifi Co , the air dens on in the heatin the number of t and the calcula 08kg/s, the hea .2kw/Co and th water and back- er is Twin-Twout= e process, heat wing: K1= -19.3 eters of air-con ronic Engineeri of the FTC odel based on f shown as gure (6). agram of propose structure xchanger repre represents G epresent the tional controll variant gain f ements the com control. The PI nd reject the d , reject the n and reduce the dness of the pro HVAC system he HVAC syste s 10 m length, ic heat capacity sity is 1.2 kg/m ng, ventilation taking a breath ation of the sup at resistance of he temperature -water to the he = - 5Co . At last exchanger, are 35Co . s /kg, T1= nditioning room ing (CIM09) fault tolerant c a block di ed model based F esents G1(s), th G2(s), NNM ( plant model, ler and the s for the sensor mbination of d ID control can disturbance, qu noise, overcom overshoot. oposed techniq m, the simulatio em: the volume 8m width and y is Ca=1.0 m3 , Based on th and air-condit h in air-conditio pply air is f wall is error of supply eat exchanger t the parameter e calculated as =30s . The m are calculate control iagram FTC he air- (neural , PID ensors r. The double n make uicken me on ques, n is e of air 4.5m he ioning oning y in rs of ed as 607
- 5. The Sec Malacca following: object G(s) and the be express Simula FTC and validate th control in t K(s) to 0.4 result. It is don’t effe there’s ef controller the simul temperatur Figure 7 Simula FTC and certify the the contro simulate w gain of K1K2=9.28 T2=406s, w object G(s) G s cond Internatio a, Malaysia, Ju K2=0.4Co s /k ) can express a . , . disturbances ed as: 0.052 0.81 0. tion (1) Senso PID Controlle he rejection to the HVAC sys 4 i.e. sensor fau s seen that the ected with d ffectiveness o (Note. The res lation have re at 20 Co i.e. 7 Sensors deterio tion (2) Param PID Controlle robust of FT olled process while the contro controlled pr 8 and the time which both in ) can express a . onal Conferenc une 2-3, 2009 kg, T2=338s. So as: [2] . ................. from outdoor 926 or Deterioratio er with norma o the sensor de tem we change ult = 0.4 figur e output of FT deterioration s on the system sults of the res been normal 4 mV indicate oration effectiven meter Variatio er with norma control in the parameters a ollers are kept rocess, K1K2= e constant, T2= ncrease 20%. as: , ce on Control, I o the controlled .........(8) ..............(9) and indoors ca on Rejection f al behavior. T eteriorate of F e the gain sens e 7 illustrate th T control syste sensor where m in the PI sponse shown lized to roo s 20 Co ). ness on the syste on Rejection f al behavior. T HVAC system are changed unchanged. Th =7.74, becom =338s, becom The controlle . Instrumentation d an for To FT or he em as ID in om em for To m, to he mes mes ed The Fi Sim Contro that t effecti system Fig Sim PID C rejectio HVAC proces contro (1). Fi nomin Fig. 10 6- In been p and pla n and Mechatr e simulation re igure 8 paramete mulation (3) N oller with norm the controller veness while P m. gure 9 the noise e mulation (4) D Controller with on to the dis C system, the ss increases 0.1 llers paramete ig. 10 show FT al PID and rob 0 comparison of Conclusion this paper, a proposed based ant parameter ronic Engineeri sults are shown er variation and t Noise Rejectio mal behavior. r FTC remai PID lose posse effectiveness on Disturbance Rej normal behav sturbances of input disturb 1 in ramp at 0 rs are kept as TC with distur bust. disturbance effe nominal PID fault-tolerant d on rejection variation. This ing (CIM09) n in figure 8. the robust of FT n for FTC an Figure (9) is s ins robust w ess of the cont n the PID control ejection for FT vior. To valida FT control i bance of cont 0th second whi same as Simu rbance is faste ects on the system control schem sensor deterio s controller des TC d PID shown without trol on ller TC and ate the in the trolled ile the ulation er than m with me has oration sign is 608
- 6. The Second International Conference on Control, Instrumentation and Mechatronic Engineering (CIM09) Malacca, Malaysia, June 2-3, 2009 compatible with HVAC systems where the traditional PID controller unable to deal with plant has nonlinear, pure lag and high inertia that are presented in the HVAC system. Thereby FTC more efficient than PID; Increased efficiency tends to reduce the cost associated with operating the HVAC system. 7. References [1] Silva, P.M.; Becerra, V.M.; Khoo, I.; Calado, J.M.F (2006) “Multiple-Model Fault Tolerant Control of Terminal Units of HVAC Systems” International Symposium on Industrial Electronics, Volume 4, 9 July 2006 Page(s):2896 - 2901 (2006) [2] Wang, Jiangjiang; Zhang, Chunfa; Jing, Youyin; (2007)”Hybrid CMAC-PID Controller in Heating Ventilating and Air-Conditioning System” International Conference on Mechatronics and Automation, Pp.3706 – 3711, 5-8 Aug. 2007 [3] Omid Omidvar ; david L. Elliott (1997)”Neural systems for Control” Chp. 5 Pp.93-101 February (1997) [4] C. Li, W. Zhang and Y. Liang, “The Decoupling control of constant temperature and humidity air- conditioning system,” Refrigeration and Air-conditioning”, vol. 6, no. 4, pp 42-47, August 2006. [5] Zdzislaw Bubnicki (2005) “Modern Control Theory” [6] (L.R. Medsker and L.C. Jain 2001) ” Recurrent Neural Networks” [7] JR Leigh (2004) “Control Theory” Second Edition chp.17 Pp.225-228 (2004) [8] Haider Abbas Fadhl Mohamed (1998) thesis “Application of Neural Network for Modeling and Control in Induction Machine” [9] Y. Zhu (2001)” Multivariable System Identification for Process Control” October 2001, Publisher: Elsevier Science & Technology Books [10] Paul S,erban Agachi, Zolt_n K. Nagy, Mircea Vasile Cristea, and Arpad Imre-Lucaci (2006) “Model Based Control” Chp. 1 Pp.14 publisher WILEY-VCH Verlag GmbH & Co. KGaA 8. BIOGRAPHIES Mr. Raad Z. Homod He graduated with a B.Eng. in Mechanical Engineering from the University of Basrah, Iraq in 1991. He worked as engineer for five years in General Establishment of Steel & Iron (Iraq) and then he worked ten years as HVAC engineering in Al-Tomoh Al-Kabir Company (Libya). Currently, he is perusing his Master in HVAC Systems and Control at the Department of Mechanical Engineering in University of Malaya. DR. T.M. Indra Mahlia received B.Eng from the University of Syiah Kuala, Indonesia, and M.Eng.Sc and PhD from the University of Malaya. He is currently an Associate Professor at the Department of Mechanical Engineering, University of Malaya, Kuala Lumpur, Malaysia Dr. Haider A. F. Mohamed received his PhD in Electrical Engineering from the University of Malaya, Malaysia in 2006. He worked as a computer engineer for two years and as a researcher for four years before he became a lecturer in the Department of Electrical Engineering in University of Malaya, Malaysia, in 2000. His main research fields are identification and nonlinear intelligent control of various systems such as robot arms, automated guided vehicles, and electric drives. 609