ѧѧ
ACTASCIENTIAECIRCUMSTANTIAE
19985µ185ڿƼڿ
˹ڵ·ͨԤеӦ
( żƼ*)
((㽭ʡԺݡ310007))
(˶)
((㽭ѧϵݡ310027))

ժҪ˹кǿԤ⹦.ĳ·ͨʵݣӦ˹е·ͨԤ⣬飬ֵʵֵӽӶΪ·ͨԤṩһµ;.
ؼʡ磻ͨԤ⣻򴫲㷨.
AN APPLICATION OF ARTIFICIAL NEURAL NETWORK TO TRAFFIC NOISE PREDICTION

Zhang Jiping
(Department of Civil Engineering, Zhejiang University, Hangzhou 310027;
Zhejiang Research and Design Institute of Environmental Protection. Hangzhou 310007)
Wu Shuoxian
(Department of Architecture, Zhejiang University, Hangzhou 310027)

ABSTRACTArtificial neural network is powerful in prediction. Based on the traffic noise data measured at a highway, the traffic noise levels were predicted with an artificial neural network. It is proved that the predicted data are in agrement with measured data well. Therefore, artificial neural network provides a new way for traffic noise prediction.
Keywordsneural network, traffic noise prediction, back-propagation algorithm.
1
·ͨԤĿǰϽϻԾĻѧо.ԻȾơ滮ȷҪ.Ϊֹĵ·ͨԤⷽҪɹΪ.һǸ봫ԭۼԤⷽ12.ʵʽܱͨ߻ȽϸӣһĲ෽ͨ½Ϊӵıʽʵһ.һǸѧͳƷԤⷽ34ݣӦûعóĵ·Ԥ⹫ʽ.෽ҪϵľϢӦڷǲ͵⽻ͨ⸴ӵܱ߻.ͨ๫ʽӦһľ.Ȼ෽ȱ.һ෽ͼۺ෽ŵ56ݣʽܸӣвʵӦ.ĿǰԤⷽھӦиļֵ.ĽӦ˵˹䲢չһ෽.
˹()о൱ԾģʽʶŻ㡢ϢܻӿƺźŴȷõ㷺Ӧã7.ڻѧ򣬰·ͨԤ⣬δӦõ.ĸĳ·ͨʵݣԤ·ͨģ.

2˹
һģϵͳܵĽƷһȺ˹Ԫͨͬģʽ.Ŀǰѷչ8.УԶǰΪ죬ӦҲΪ㷺.
ǰ㡢.һ.ÿɸ˹Ԫɣ˳ӣͼ1.
kеjԪΪy(k-1)i(i=1,2,,Nk-1)Ϊy(k)j.ĹϵΪ

(1)
(j=1,2,,Nk; k=1,2,M)
ʽУw(k-1)ijΪk-1еiԪkjԪȨֵ(k)jΪӦԪֵf(k)j()ΪԪĴݺNkΪkԪĿMΪܲ.


ͼ1͵˹
Fig.1 A typical artificial neural network
ݺ׼ԺɵȡΪʽļѹ(Sigmoidal funtion):
f(x)=A/1+exp-(B*x+)ݣ
(2)
ʽУxΪԪ룬ΪֵA,BֱΪֵбʵڲ.

Ӧݶ½ԭͨźŷ򴫲㷨(BP㷨)֪롢ѵԵźţԶѧϰڲȨֵWijֵjı9.
ѵõĶǰԤģ͵Ĺ.ɽǰһ䡱ʵõΪ롰䡱ѧϰ.Ӱڶѧϰɡ䡱Զ¼.ڽڵԪݺǷԵģˡ䡱Ҳз.ѧϰ̾Ԥģ͵Ḷֻ̌Ҫڵѵ㹻࣬䡱ʵֶԤ.

3·ͨԤ
ͼ1һ3ԪǰΪ·ͳĳٺ뿪·룬Ϊ㴦ĵЧLeq.ڵ㡢ڵ㡢ڵֱΪ15131.


ͼ2ʾͼ
Fig.2 Section plan of the measurement
ѵѧϰͼΪ320ߺϺK217+200·ͨʵ.ʱ4̨HS6211ͻԶͳƷǣڹ·һչڷֱ·紹ֱ204060100mһ㣬ͬ.ͼ2ʾ.17ʱΣÿ20min10minաл(GB3222-82)͡򻷾(GB/T14623-93)ейؽͨ淶Уͬʱ¼ͳĳƽ.ֳСγСͳͿͿͳĦг7.ÿβӡ¼Ч.ڳͳʱ仯˳βΪѧϰѵ磻żβΪԼǰ罻ͨԤģ.


ͼ3ѵѧϰ̿ͼ
Fig.3 Flow diagram of the training and learning
ӦBP㷨ѵ·ͨԤģ.Ӧѵ̼ͼ3.岽Ϊ
(1)ӦڵԪȨֵʼ.
(2)ѵѡȡһѵԣһy(0)j(j=1,2,,15)һĿd1.y(0)1y(0)2y(0)15αʾʵʱڸֵĳ䳵٣Լ뿪·ľ룻d1ʾӦʱβͬʵĵЧ.
(3) ѡȡˣ㰴(1)ʽ(2)ʽм㣬õy(3)1.y(3)1ʾóĵЧ.
(4)ƫֵS=(d1-y(3)1)2.
(5)ݲֵݶ½ԭʽȨӺֵ(Ƶԣɲοף10)

(3)

(4)
ʽУΪѵϵͬʽ(1)ʽ(2)k=1ʱi=1,2,,15k=2ʱi=1,2,,13,k=3ʱi=1.

(6)ѵѡȡ2ѵԣظ(3)(4)(5)(6)裬õƫֵ.
(7)ѵѵԶʹùһκһѵڣƫƽͣ

(ѵ9ʱʵУ35εĲЧ).
(8)صϵ(2)裬һѵڣõһѵڵĽ.ظѭõS(M)ֱS(M)Сĳ涨ֵ.
ѵDX486/66PC16λCƽִ̨Уȡ0.1Լ107min72074ѵѭƽÿԼ11ѭ0.001dB(A)ľֵ.㡢м㡢Ȩֵֵѧϰֱڱ14.
1ȨWijѵ
Table 1 Training results of the weight wij in the neural network from input line to input layer

Wij
j:ڵ(мڵ)
123456789101112131415
i(ڵ)1-0.908-0.363+0.196-0.149+1.704-1.012+0.126+1.236+1.759+0.514-3.578-2.045-1.383-1.202+1.852
2-1.463-0.251+0.286+0.291+1.124+3.334-1.788-1.931-3.674-2.117+1.726+2.616+3.379+3.783+3.249
3-0.078+2.213+0.016+0.478+1.792-0.478-1.898+3.743-3.404-2.047+0.308+0.337-2.151-0.780+0.329
4+1.768+0.501-0.534-0.625+2.400-0.737-1.087+2.498-1.442-2.104+0.917-2.825+0.257-1.581+0.269
5-2.397-1.685+0.793+1.174+1.938-2.030+0.013+1.430-2.367-0.643+2.925+3.869-0.819-3.954-0.832
6+0.376+0.432+0.513-0.425+0.302-2.995-0.667-0.065+0.537-0.062+1.576+2.105+2.255+1.981+2.020
7+0.835+2.102+0.754-0.326+1.459-0.313-2.175+2.101-2.661-2.960-0.411-1.634-2.435-0.291-0.218
8-4.774-3.263-0.880-0.515-1.134-3.059+1.223-0.017-1.407+2.578+0.453+3.119+1.399+1.294-3.472
9-1.137-1.180+0.260-0.091+0.740+0.071-1.008-0.510-0.924+1.578+0.227-0.027+0.737+0.592-1.104
10-0.928-1.374+0.465-0.217-0.895+0.738-0.077+2.149+0.963+0.686+0.030-1.079-0.909+1.109-0.888
11-1.648-0.848-0.143+0.334-0.723-1.001+0.308+2.101+0.743+1.144-0.565-1.252+0.339+0.362-2.172
12-2.723-1.954-0.405+0.077-0.949+0.107+3.157+0.937+0.147+2.337+2.452+1.043+0.295+0.287-2.364
13-0.988-0.868+0.415+0.577-0.841+0.167+1.625+1.332+0.0012.477+1.048+1.152+0.133+0.333-1.545
14-0.757-2.153-0.269+0.372-0.051+0.293+0.897+0.556+0.210+1.465+1.021+1.117+0.809-0.160-2.610
15+5.099+2.946+0.398+0.135-0.161-0.656+0.210+2.575-0.770-0.938+0.742+1.612+0.808-0.318+0.145

2мȨWijѵ
Table 2 Training results of the weight wij in the neural network from input layer to middle layer

Wij 
j:мڵ(ڵ)
12345678910111213
i:ڵ(мڵ)1-0.882-1.522-2.330-1.725-1.564-2.239-3.523-1.997-2.893-2.925-2.648-1.227-2.845
2-1.445-1.121-2.326-2.005-3.197-4.249-4.371-2.490-2.705-2.666-3.437-2.920-3.150
3-0.407+0.377-0.923-1.393-1.093-2.360-2.206-2.656-1.009-0.893-2.406-1.261-0.771
4+0.756-0.409-0.732-1.012-2.147-0.579-1.350-0.962-1.137-1.706-1.000-0.159-0.965
5-1.614+0.293+0.340+0.593+0.271+1.152+1.167+2.513+3.070+4.167+3.305+3.824+5.348
6-0.174-1.346-2.845-2.623-3.252-3.598-3.120-3.271-1.793-2.679-3.311-3.100-3.256
7-0.828-1.509-1.636-1.109-1.485-1.659-3.117-1.695-1.790-2.575-1.200-1.788-2.102
8-1.226+0.661-0.039-0.313+0.648+0.559+2.128+2.283+3.265+2.870+4.348+4.461+4.965
9-0.437-0.952-2.069-1.313-2.704-2.723-2.663-2.746-2.589-1.791-2.838-2.398-2.129
10+0.498+0.144-1.827-2.327-2.685-1.674-2.151-2.039-2.643-3.576-1.841-2.904-3.171
11-1.855-0.986-2.172-1.313-1.810-2.127-1.839-2.089-2.297-1.110-1.830-1.707-1.362
12-1.273-0.603-1.023-1.169-1.810-2.408-3.225-1.747-2.959-1.301-3.023-1.226-2.351
13+0.010-1.841-1.075-2.671-2.982-2.147-3.538-3.564-2.769-2.200-2.484-1.661-1.575
14-1.192-1.412-2.334-1.869-2.726-3.236-2.573-2.066-3.269-3.534-3.231-2.081-2.927
15-1.201-1.029-1.127-1.560-2.969-1.752-2.192-2.995-2.671-1.669-2.765-1.947-2.180

3мȨWijѵ
Table 3 Training results of the weight wij in the neural network from middle layer to output layer

Wij
iмڵ(ڵ)
12345678910111213
j(ڵ)-5.631-2.830+1.556+4.484+8.488+12.845+17.567+19.191+22.498+25.648+29.360+30.508+34.592

4ڵԪֵѵ

Table 4 Training results of the neurode thresholds in each layer of the neural network

Ԫ
(ڵ)123456789101112131415
+0.173-0.003+0.269-0.184+0.154+0.615-0.626+0.678-0.600+0.545-0.894-0.073-0.780+0.244-0.171
м+0.214+0.376+0.381-0.162+0.543-1.039+0.807+0.114+1.117+0.238+0.879+1.173+2.387
-12.527


ѵõе·ͨԤ⣬ֵʵֵ(ѧϰͼ)ıȽϼ5.
5ԤʵֵȽ
Table 5 Comparison between predicions and measurements in the neural network

*
/hLeqdB(A)**
/hLeqdB(A)
20m40m60m20m40m60m100m
1786ֵ
ʵֵ67.8
67.864.7
64.762.0
62.056.0
56.02726ֵ
ʵֵ
ֵ71.2
69.3
+1.969.6
67.4
+2.265.9
64.4
+1.556.5
57.1
-0.6
3486ֵ
ʵֵ70.5
70.567.1
67.163.2
63.256.7
56.74669ֵ
ʵֵ
ֵ69.4
69.1
+0.364.7
65.8
-1.161.9
61.1
+0.856.3
54.5
+1.8
5612ֵ
ʵֵ70.9
70.967.1
67.163.4
63.458.1
58.16546ֵ
ʵֵ
ֵ68.5
68.7
-0.265.2
65.6
-0.462.9
62.0
+0.957.3
55.1
+2.2
7726ֵ
ʵֵ70.2
70.265.9
65.963.9
63.960.3
60.38504ֵ
ʵֵ
ֵ72.3
69.4
+2.970.6
66.1
+4.565.1
62.0
+3.154.6
54.0
+0.6
9678ֵ
ʵֵ68.3
68.365.7
65.763.6
63.660.8
60.810708ֵ
ʵֵ
ֵ72.5
68.9
+3.672.2
66.5
+5.768.5
64.0
+4.559.4
59.8
-0.4
11531ֵ
ʵֵ68.5
68.566.3
64.5
64.561.7
61.712420ֵ
ʵֵ
ֵ64.9
67.3
-2.463.0
65.0
-2.061.8
62.3
-0.559.7


13246ֵ
ʵֵ64.6
64.663.0
63.061.9
61.959.5
59.514324ֵ
ʵֵ
ֵ63.7

62.1
62.7
-0.661.3
61.4
-0.159.2
58.9
+0.3
15414ֵ
ʵֵ66.4
66.464.3
64.362.8
62.860.6
60.616258ֵ
ʵֵ
ֵ63.5
64.8
-1.362.7
62.6
+0.161.0
61.2
-0.258.4
58.5
-0.1
17288ֵ
ʵֵ64.4
64.463.1
63.161.5
61.559.0
59.0ע*Ϊѧϰѵ
**Ϊ


ӱ5ɼѧϰֵʵֵȫһ£ʾѵѧϰЧѧѧϰϢ֪ʶ.ѵ30ЧεļֵʵֵȽϣֵСΪ0.1dB(A)Ϊ+5.7dB(A)ƽΪ1.56dB(A)Ϊ1.48dB(A)ɴ˵ֵʵֵӽ.
4֡
BP㷨һַԵŻѧϰʱʱϳֲС⣬Ǹ÷ȱ.ǰһ⣬ѡøָĽ跨ԣͬʱ32λƽ̨486ڻϽУЧʺٶȣԺһ⣬Բ֮һǲBoltzmann10.⣬ڵĸ(رڵ)йزѡȡϵݣҪݾȷ.Ҳиֵʵֵֵƫԭ.
·ͨԤģͣԽܹͨĵ·Ƶĵ·Ԥ⳵ܱ߻·ĵЧ.и߶ȵԡģԺݴԡ֯ѧϰΪΪѵԤṩһոµֶ.еѧԤⷽȣ֪ϢĹϵ޷ͨ÷ʾʱƣһӦǰõԤⷽ.
Ԥ֮ⷽҪΪȫ桢ܴݶԣѵԺܲ.ѵݶԳʱʹһͬʱԤĹܣͬʱԤЧLeqͳL10,L50, L90
绹Ӧڻѧ棬繤ҵ豸Ԥ⣬Ԥ⣬ϸЧԤ.ЩԤкǿĿԣʺӦ缼.
* ͬʱΪ㽭ѧľϵʿ
ο
1գԭ޹֦忡ɱ.աҫ.·ʩѧ.ͨ磬1982271
2Wu Shuoxian. A Simple Method for Predicting Kerbside L10 Level from a Free Multi-type Vehicular Flow. Appl Acoustics, 1987,20:15
3Watkins L H. Environmental Impact of Roads and Traffic. England:Applied Science Publishers, 1981:36
4ƽȱ.лȾ.йѧ磬198960
5Ohta M et al. A Systematic Prediction Method for Road Traffic Noise and Ground Vibration Based on an Arbitrary Flow Model of a Non-Poisson Type. Acustica, 1991,74:217
6Zhang Jiping. A Study on the Highway Noise Prediction Model Applicable to Different Traffic Flow. Noise Control Eng J,1993,41:371
7Lisboa P G J. Neural Networks Current Applications. England:International Thomson Computer Press, 1992
8ʵȱ.Ӧü.ɳƼѧ磬1993
9ȱ.˹Ԫ.ѧ磬1994
10.ģ߼.պѧ磬1996
