http://www.paper.edu.

cn
9 2
2003 5

EL ECTROCHEM ISTR Y

Vol. 9 No. 2
May 2003

Article ID :1006- 3741- ( 2003) 02- 0228- 07

Corrosion Prediction of Carbon Steel in

Concentrated Lithium Bromide Solutions at
High Temperature Using BP Neural Net work
HUAN G Nai- bao , L IAN G Cheng- hao 3
( School of Chem ical Engi neeri n g , Dalian U ni versity of Technology , Dalian 116012 Chi na )

Abstract : Neut ral network model has been developed for t he computation of corrosion
rate of carbon steel in LiBr solutions. The result s indicate t hat t he model is capable of re2
producing t he effect s of changes in alkalinity ,temperat ures ,LiBr and Na2 MoO4 concen2
t rations on t he rates of general corrsion and good agreement between calculated and ex2
perimental corrosion rate is obtained. The model can be used to satisfactorily predict t he
corrosion rate of carbon steel in different concent rations of LiBr solutions containing dif2
ferent inhibitors at different temperat ures. It also provides a novel met hod for corrosion
monitoring of metals used in LiBr absorption chiller.

Key words : Corrosion ,Carbon steel ,Lit hium bromide ,B P neural network
CLC Number : T G 174. 42 Document Co de : A
1 Introduction
Because of a growing concern to environment welfare ,Lit hium bromide absorption chiller ,
which uses concent rated lit hium bromide as t he absorption fluid and water as t he ref rigerant ,is
gradually substit uted for ref rigerator making use of fluorohydrocarbon as ref rigerant and applied in
t he whole world owing to t heir high hydration heat ,high solubility of solid p hase ,appropriate vis2
cosity and good t hermal stability. However ,lit hium bromide solution is very corrosive toward t he
st ruct ural materials ,generally carbon steel ,used for t he solution containment part s in t he machine.
In indust ry ,t he most effective and economic met hod is to add different inhibitors to t he system ,
such as LiO H ,Li2 MoO4 ,Li2 CrO4 ,Na2 MoO4 ,LiNO3 ,PMA/ Sb ,etc. Corrosion protection properReceived date : 23 Oct . 2002

3 To whom correspondence should be addressed , Tel : ( 86- 411 ) 3631333- 3678 or 3268 , E- mail :liangch @chem.
dlut . edu. cn
project ( 972210) supported by t he natural science foundation of liaoning provine

http://www.paper.edu.cn

229

ties of t hese inhibitors have been reported in t he literat ure by several aut hors[ 15 ] . In our previous
paper , t he corrosion behavior of carbon steels and stainless steels have been st udied in different
LiBr solutions ,which contain different concent rations of LiO H and Na2 MoO4 at different tempera2
t ures under static conditions t hrough weight-loss test s and elect rochemical measurement s[ 68 ] .
In t hese days ,neural network is a rapidly growing field of artificial intelligence and has been
found many applications in t he modeling of nonlinear systems because of it s characteristic of ap2
proximating any nonlinear continuous arbit rarily well on a compact set . The most widely used
neural network is back propagation (B P) network ,which account s for most neural network appli2
cations ,for it s simplicity and it s power to ext ract usef ul information f rom samples[ 9 ] . B P neural
network has also been applied to t he st udy of corrosion behavior [ 1012 ] such as forecasting corro2
sion forms ,corrosiveness ,service life of equip ment ,etc.
A literat ure research shows no works describing t he effect of factors such as temperat ure ,
concent rations of LiBr and inhibitors ,p H etc on corrosion rate of carbon steel in lit hium bromide
absorption chiller by using neural network. This paper is a continuation of t he previous work.
Specifically ,t he st udy is expanded to develop t he model of corrosion rate of carbon steel relating to
LiBr concent rations ,inhibitor concent rations and temperat ures using B P neural network based on
t hese data we have obtained.

2 Experimental
Rectangular sheet specimen ( 30mm 20mm 2mm ) was used for corrosion loss measure2
ment and it s chemical composition is shown in Tab. 1. The specimens were ground wit h 1000-grit
SiC paper , rinsed wit h deionized water , and finally degreased wit h acetone. Test solutions were
prepared f rom a reagent grade LiBr and deionized water. Two specimens and 90cm3 of t he test so2
lution in a polytet rafluoroet hylene ( P TFE) cylinder bush ( 50mm inner diameter ,65mm lengt h) ,
which was t ransferred into a stainless steel autoclave , and t he solution was deoxygenated for 1
hour wit h nit rogen. The autoclave was t hen held at predetermined temperat ure for 200 hours in a
t hermostat ( Kosumosu A T- S13 ) . The corrosion rate was determined f rom t he weight changes of
t he specimen after t he specimen was cleaned by 3 mol/ L HC1 containing 1 % hexa- met hylenetera2
mine at 50 for 5min. The result s have been previously reported [ 68 ] .
Tab. 1 Chemical composition of t he carbon steel/ %
C

Si

Mn

Fe

0. 12

0. 01

0. 35

0. 02

0. 02

Bal

3 BP net work algorithm

3. 1 Principle of BP net work algorithm
B P neural network is basically a gradient decent algorit hm designed to minimize t he error

http://www.paper.edu.cn

230

2003

f unction in t he weight s space. A typical

t hree-layer B P neural network is shown
in Fig. 1. Commonly used activation f unc2
tion is t he Sigmoid f unction. During
t raining of t he neural network , weight s
and t hreshold values are adjusted to de2
crease t he total error ,which is defined by
t he following equation ,so as to make t he
calculated output data as close to t he giv2
en output data as possible.

Fig. 1 Structure diagram of a t hree-layer BP neural network

1
( T pk - Y pk) 2
2 p k
Where P indexes t he patterns in t he t raining set ,

E = Ep =
p

indexes t he output unit s , T is t he act ual

value of output layer , Y is t he calculated value of output layer.

Based on t he input of t he input layer , modification formulas of weight s can be obtained by
adopting gradient descent met hods.

k bj
V jk = - ( 5 E/ 5 V jk ) =
W ij = - ( 5 E/ 5 V jk ) = ej X i
Where X i is t he input of input layer , bj is t he output of t he hidden layer , ej is t he error of t he
hidden layer ,k is t he error of t he output layer , is t he step size of t he st udy ,W ij and V jk are
t he weight changes.
In order to accelerating convergence rate and avoiding oscillating , Hoskine [ 13 ] gave momen2
t um st rategy met hods by adding moment um st rategy to weight s updating rule. The new equations
are given by :

V jk ( n )
V jk ( n + 1) = - ( 5 E/ 5 V jk ) +
W ij ( n )
W ij ( n + 1) = - ( 5 E/ 5 V jk ) +
Where is t he moment um coefficient , n is t he cycle index.
In t he B P algorit hm ,t raining is carried out by repeatedly presenting t he entire set of t raining
patterns until t he error f unctions over all t he t rainings are minimized and wit hin t he preset t rain2
ing precision.

3. 2 Tra ining BP neural net work

3. 2. 1 Data initialization of B P neural network
Lit hium bromide concent rations between 55 % and 65 % are commonly used in absorption
chiller. This range provides maximum efficiency for absorption of t he flash-evaporated water.
Working temperat ure varies between 145 and 173 since LiBr concent ration ranges f rom
55 % to 65 % in t he absorption system. Recently , t he t riple-effect absorption cycle has att racted
much interest to replace t he conventional machine as t he more efficient one. In high generator ,t he

http://www.paper.edu.cn
231

temperat ure of lit hium bromide solution is raised to 200 . In indust ry ,LiO H and Na2 MoO4 are
t he popular inhibitors.
In principle ,it has been proved t hat any continuous f unction can be uniformly approximated
by a neural network model wit h only one hidden layer. So a t hree layer B P network is employed in
our st udy. It is easy to determine t he number of neurons in t he input layer and t he output layer
during applications. There are four neurons in t he input layer denoting temperat ure and t he con2
tent of LiBr ,LiO H ,Na2 MoO4 in absorption chiller ,respectively. However it is not easy to choose
t he appropriate number of neurons in t he hidden layer for t here is currently no definite rule to de2
termine it . Using too many neurons impedes generalization and increases t raining time. Using too
few neurons impairs t he neural network and prevent s t he correct mapping of input to output . In
t his st udy ,t he number of neurons in t he hidden layer is determined according an empirical formu2
la [ 14 ] and is eight . There is only one neuron in t he output layer representing t he corrosion rate.
Data of 79 typical corrosion rate samples are gat hered for use in t raining and testing t he B P
network. The number of corrosion rate in 55 % LiBr solution is 20. 30 patterns are t he data in
60 % LiBr solution. The rest patterns are t he data in 65 % LiBr solution. The database is divided
into two data set s :t raining set s and test set s ,containing 88 % and 12 % data of database ,respec2
tively. Preprocessing of t he data is usually required before presenting t he patterns to B P neural
network. It is necessary because t he Sigmoid activation f unction modulates t he output of each neu2
ron to values between 0 and 1. The following normalization procedure is commonly adopted and is
used in our work.
X max - X i
1

2 X max - X min
Where X max and X min are t he maximum and minimum values of a variable , X i is t he normal2

X i =

ized value of t he variable.

Initial weight s and t hreshold values are randomly initialised wit h a uniform dist ribution over
[ - 1 ,1 ] . Step size and moment um coefficient are chosen to bo 0. 1 and 0. 9 ,respectively. During
t he t raining of t he B P network ,t he error f unction for all t raining patterns is monitored. The t rain2
ing course is stopped when t he error f unction reduces wit hin a given tolerance. And t hen t he fixed
st ruct ure of B P network is obtained.
3. 2. 2 Validating B P neural network
Of t he total t raining patterns ,10 are randomly chosen and are used to validate whet her t he
t rained B P network has mapped t he nonlinear relationship of t hese factors. The tested result s are
shown in Tab. 2. As shown ,t he relative errors of predicted values of corrosion rate are wit hin
10 % ,and t he maximum absolute error is 8. The predicted result s agree wit h t he act ual data to
wit hin reasonable experimental error. Thus ,t he result s indicate t hat t he t rained B P network esti2
mates corrosion rate wit h and accuracy similar to t hat in t he experimental measurement and can be
used to predict t he corrosion behavior of carbon steel.

http://www.paper.edu.cn

232

2003

Tab. 2 Results of validating BP network using training samples

LiBr

Temperature

Concn. / %

LiOH Concn. Na2 MoO4 Concn. Experimental

/ mol
L-

mg
L

- 1

data/my -

Absolute

Predicted

Relative

value/my - 1 error/my -

error/ %

145

150

536. 4366

535. 0343

1. 4023

0. 26

55

145

0. 07

150

36. 6015

34. 0258

2. 5757

7. 04

55

175

0. 07

150

49. 0591

49. 6773

0. 6182

1. 26

60

145

0. 07

150

39. 5889

39. 6806

0. 0917

0. 23

60

150

0. 02

200

364. 7655

372. 6786

7. 9131

2. 17

60

150

0. 10

150

40. 6374

37. 0702

3. 5672

8. 78

60

200

0. 07

150

121. 7986

125. 8548

4. 0562

3. 33

65

173

0. 07

1000

51. 3772

50. 5059

0. 8713

1. 70

65

173

0. 10

406. 4175

407. 2542

0. 8367

0. 21

65

240

0. 10

300

373. 5200

368. 5536

4. 9663

1. 33

3. 3 Predicted results of BP net work and discussion

The predicted values of t he test set s ,which are obtained wit h t rained B P network ,are shown
in Tab. 3. The relative errors of t he ot hers are wit hin 5 % except t hat one pattern has larger rela2
tive errors in 55 %LiBr solution ,which is 18. 30 %. The reason is t hat t he amount of corrosion data
in 55 %LiBr in input patterns is less t han t hat in ot her cases ,resulting in higher relative errors in
less concent rated LiBr solutions. So far as absolute errors are concerned ,t he absolute error of t he
pattern having large relative error is less t han 7. The accuracy still meet s t he requirement . The re2
sult s indicate t hat t rained B P network can be used to predict t he corrosion behavior of carbon steel
in different concent rations of LiBr solutions wit h various inhibitors at different temperat ures.
Tab. 3 Predicted results of trained BP network
LiOH Concn. Na2 MoO4 Concn. Experimental

LiBr

Temperature

Concn. / %

55

145

55

160

0. 07

55

200

60
60

/ mol
L

- 1

- 1

mg
L
150

data/my

- 1

Predicted
value/my

Absolute
- 1

error/my

Relative
- 1

error/ %

36. 6100

29. 9105

6. 6995

18. 03

150

41. 0428

39. 6497

1. 3931

3. 39

0. 07

150

84. 8376

81. 0565

3. 7811

4. 46

150

0. 02

200

364. 7655

372. 6786

7. 9131

2. 17

150

0. 07

150

42. 3084

44. 4279

2. 1195

5. 01

60

240

0. 07

150

242. 5900

240. 2126

2. 3774

0. 98

65

173

0. 10

300

74. 6300

72. 7456

1. 8844

2. 53

65

173

0. 15

150

302. 0973

291. 4716

10. 6257

3. 52

65

220

0. 07

150

444. 4600

466. 5221

22. 0621

4. 96

http://www.paper.edu.cn

233

4 Conclusions
B P neural network was t rained using obtained corrosion data and was tested wit h t he remain2
ing patterns. The predicted values agree wit h experimental data. The t rained B P network can be
used to predict corrosion rate of carbon steel in LiBr solutions having different concent rations wit h
different inhibitors at different temperat ures. Predicted accuracy of neural network relates to t he
selection and t he number of t raining patterns. The model provides a novel met hod for st udying t he
corrosion of metallic materials and field monitoring in LiBr absorption chiller. .

Reference s :
[1 ] Kazuo Tanno ,Shiro Saaito ,Masahiko Ito ,et al. Corrosion inhibition of carbon steel ,copper and 70/ 30 cupronickel by benzotriazole in concentrated lit hium bromide solution at elevated temperature [J ] . Boshoku Gijtsu ,
1984 ,33 :516519.
[ 2 ] Masahiko Itoh ,Akira Minato ,Michihiko Aizawa ,et al. Corrosion inhibition of carbon steel and cupro- nickel by
lit hium molybdate in concentrated lit hium bromide solution at elevated temperature[J ] . Boshoku Gijtsu ,1984 ,
33 :504508.
[3 ] Tanno K , Itoh M , Takahashi T ,et al. The corrosion of carbon steel in lit hium bromide solution at moderate
temperatures[J ] . Corrosion Science ,1993 ,34 ( 9) :14411 451.
[4 ] Tanno K ,M. Iton M , Hekiya H ,et al. The corrosion inhibition of carbon steel in lit hium bromide solution by
hydroxide and molybdate at moderate temperatures[J ] . Corrosion Science ,1993 ,34 ( 9) :1 4531 461.
[ 5 ] Hitoshi Yashiro ,Akijiko Sai ,Naoaki Kumagai ,et al. Corrosion behavior of carbon steel in concentrated LiBr so2
lutions up to 473 K[J ] . Zairoy- to- Kankoy ,1999. . 48 ( 6) :369374.
[ 6 ] Guo Jian- wei ,Liang Cheng- hao. Mechanism of t he formation of passive state films on carbon steel in a high
temperature and high concentration lit hium bromide solution [J ] . Materials Science and Technology ( in Chi2
nese) ,2001 ,9 ( 4) :413416. .
[ 7 ] Guo Jian- wei ,Liang Cheng- hao. Corrosion behavior of carbon steel in high temperature and high concentration
lit hium bromide solution[J ] . J . of Dalian University of Technology (in Chinese) ,2000 ,40 ( 5) :554556.
[ 8 ] Guo Jian- wei. Research of inhibition mechanism on absorption lit hium bromide refrigerator[ D ] . Dalian :Dalian
University of Technology ,2000.
[ 9 ] Chun- gen Yin ,Zhong- yang L uo ,Mingjiang Ni ,et al. Predicting coal ash fusion temperature wit h a back-propa2
gation neural network model[J ] . Fuel ,1998 ,77 ( 15) :1 7771 782.
[ 10 ] Smets H M G ,Bogaerts W F L . SCC analysis of austienitic stainless in chloride- bearing water by neural net2
work techniques[J ] . Corrosion ,1992 ,48 ( 8) :618623.
[ 11 ] Rosen E M ,Silverman D L . Corrosion prediction from polarization scans using an artificial neural network in2
tegrated wit h an expert system[J ] . Corrosion ,1992 ,48 ( 9) :734744.
[12 ] Trasatti S P ,Mazza F. Crevice corrosion ,a neural network approach [J ] . British Corrosion Journal ,1996 ,31
( 2) :105112.
[ 13 ] Hoskins J C , Himmelblau D M. Process control via artificial neural networks and reinforcement learning[J ] .
Comput . Chem. Eng. ,1992 ,16 ( 4) 141151.
[14 ] Mirchadani G ,Cao W. On hidden units in neural nets[J ] . IEEE Trans Circuits Syst ,1989 ,36 ( 15 ) : 661
664.


234

http://www.paper.edu.cn

2003

B P
, 3
( , 116012)

: LiBr . , ,
LiBr Na2 MoO4 , ,

LiBr , ,
.

: ; ;LiBr ;BP