Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013 1591<br />
∑<br />
neth = w _ xhih⋅X i<br />
−u _ hh<br />
⎫<br />
⎪<br />
i<br />
⎬<br />
Hh = ( neth) = 1 1+ exp( −neth)<br />
⎪⎭<br />
Comput<strong>in</strong>g output vector Y of output layers:<br />
net<br />
j<br />
= ∑ w _ hyhj ⋅Hh −u _ y ⎫<br />
j<br />
⎪<br />
h<br />
⎬<br />
Yj = ( netj) = 1 (1 + exp( −netj))<br />
⎪ ⎭<br />
4) Comput<strong>in</strong>g the difference mountδ<br />
δ<br />
j<br />
(18)<br />
(19)<br />
Comput<strong>in</strong>g difference mount of output layers<br />
δ = Y (1 −Y )( T − Y ) (20)<br />
j j j j j<br />
δ<br />
Comput<strong>in</strong>g difference mount of hidden layers<br />
h :<br />
δ = H (1 −H ) ∑ w_<br />
hy δ (21)<br />
h h h hj j<br />
j<br />
5) Comput<strong>in</strong>g correction mount of weigh dw , and<br />
correction mount of threshold du .<br />
Comput<strong>in</strong>g weigh correction mount of output layers<br />
dw _ hy , correction mount of threshold du _ y :<br />
dw _ hyhj = δ<br />
jH<br />
h ⎫⎪ ⎬ (22)<br />
du _ y<br />
j<br />
=−ηδ<br />
j ⎪⎭<br />
Comput<strong>in</strong>g weigh correction mount of hidden layers<br />
dw , correction mount of threshold du :<br />
dw _ xhih = ηδh X<br />
i ⎫<br />
⎬ (23)<br />
du _ H<br />
h<br />
=−ηδh<br />
⎭<br />
6) Updat<strong>in</strong>g weigh mount w_<br />
hy , and threshold<br />
u_<br />
y.<br />
Updat<strong>in</strong>g weigh mount of output layer w_<br />
hy and<br />
threshold u_<br />
y:<br />
w _ hyhj = w _ hyhj + dw _ hyhj<br />
⎫⎪ ⎬ (24)<br />
u_ yj = u_ yj + du_<br />
yj<br />
⎪⎭<br />
Updat<strong>in</strong>g weigh mount of hidden layer w_<br />
hy and<br />
threshold u_<br />
y:<br />
w _ xhih = w _ xhih + dw _ xhih<br />
⎫<br />
⎬ (25)<br />
u_ hh = u_ hh + du_<br />
hh<br />
⎭<br />
(c) The test<strong>in</strong>g of BP neural network<br />
After be<strong>in</strong>g build<strong>in</strong>g, the character of model must be<br />
tested by us<strong>in</strong>g the samples which are not used <strong>in</strong><br />
build<strong>in</strong>g the model, so that the Correctness and<br />
Practicality of the model can be verified.<br />
The comput<strong>in</strong>g format of the test<strong>in</strong>g process is showed<br />
as followed.<br />
1) Adopted the stable Weight matrices after tra<strong>in</strong>ed<br />
w_<br />
xh, w_<br />
hy and Threshold vector u_<br />
h, u_<br />
y.<br />
2)Input vector X of test<strong>in</strong>g samples.<br />
3)Comput<strong>in</strong>g output vector Y .<br />
Comput<strong>in</strong>g the output vector of hidden layer H :<br />
∑<br />
neth = w _ xhih⋅X i<br />
−u _ hh⎫<br />
i<br />
⎪<br />
1 ⎬<br />
Hh = f( neth)<br />
= ⎪<br />
−neth<br />
1+ exp ⎪⎭<br />
(27)<br />
Comput<strong>in</strong>g the output vector of hidden layer Y :<br />
net<br />
j<br />
= ∑ w _ hyhj⋅H h<br />
−u _ y ⎫<br />
j<br />
h<br />
⎪<br />
1 ⎬ (28)<br />
Yj = f( netj)<br />
=<br />
−net<br />
⎪<br />
j<br />
1+ exp ⎪⎭<br />
V. THE EXPERIMENT AND ANALYSIS OF BP NEURAL<br />
NETWORK FOR WATER QUALITY ASSESSMENT<br />
In the learn<strong>in</strong>g process of network, some standard<br />
water quality classification is adopted <strong>in</strong> learn<strong>in</strong>g samples.<br />
With consider<strong>in</strong>g that the range of activation function<br />
is[0,1] , and water quality classification is from the first<br />
class to the fifth class, so the five water quality<br />
classifications are only part of the whole range, and no<br />
attach<strong>in</strong>g the limited values 0 and 1.<br />
In this paper, target outputs are 0.1,0.3,0.5,0.7,0.9 ,<br />
and the output represents No.1-5 water quality<br />
classifications. As the most important parameters <strong>in</strong><br />
debugg<strong>in</strong>g the BP network, learn<strong>in</strong>g rate η = 0.68 ,<br />
Impulse coefficient α = 0.5 , then the network can be<br />
tra<strong>in</strong>ed after 1600 iterations.<br />
(a) Errors curve of learn<strong>in</strong>g process<br />
(b) Test<strong>in</strong>g result curve of network for some samples<br />
Fig. 4 The learn<strong>in</strong>g process curve of the network<br />
© 2013 ACADEMY PUBLISHER