Professional Documents
Culture Documents
net/publication/226900543
CITATIONS READS
4 603
3 authors, including:
Shuyong Jiang
Harbin Engineering University
89 PUBLICATIONS 743 CITATIONS
SEE PROFILE
All content following this page was uploaded by Shuyong Jiang on 21 January 2016.
DOI 10.1007/s11595-009-5791-0
Abstract: 50CrVA cold-rolled spring steel strip was used to fabricate the diaphragm of the
automotive horn. The material parameters which were taken into account in the design of the dia-
phragm involve elongation, elastic limit, Young’s modulus, yield strength and tensile strength. The
tempering process was carried out in order to enable the diaphragm to possess the excellent mechanical
properties, such as high elastic limit, high fatigue strength and perfect stress relaxation resistance. As a
nonlinear information processing system, the backpropagation artificial neural network (BPANN) was
applied to predict and simulate the relationship between the mechanical properties of the diaphragm
and the tempering process parameters. Experimental results show that a BPANN with 3-8-5 architec-
ture is capable of predicting the relationship between the mechanical properties of the diaphragm and
the tempering temperature successfully and lays the profound foundations for optimizing the design of
the diaphragm. BPANN simulation results show that the tempering temperature ranging from 380 to
420 ℃ contributes to enhancing the comprehensive mechanical properties of the diaphragm including
high Young’s modulus, high elastic limit and high fatigue strength.
Key words: artificial neural networks; mechanical properties; tempering; diaphragm; 50CrVA
1) Initialize weights and thresholds, namely assign where Z n is the normalized value of Z , and Z max
arbitrary small random values for the weights and and Z min are the maximum and minimum values of
thresholds. Z in the database, respectively.
2) Present training data, i e, present the input vectors
and the desired outputs. 4 Determination of ANN Pa-
3) Calculate the actual output y j of the network. rameters
y j = f (∑ wij xi − θ j ) (3)
i 4.1 Input layer neuron size
where f is the sigmoid function. The number of nodes in the input layer is determined
4) Calculate the error E of the network according to the problem to be solved. As the 50CrVA
steel strip used as the diaphragm material is cold rolled,
1 2
E =
2
∑(y j
j
− y j* ) (4) the tempering heat treatment is a necessary approach to
enhance the mechanical properties of the diaphragm. The
where the jth neuron is in the output layer, y j is application of the BPANN is aimed at establishing the
the actual output and y ∗j is the target output. Assume ε relationship between the tempering parameters and the
is the allowable error accuracy, and if E < ε , then save mechanical properties of the diaphragm. The input vari-
initial weights and thresholds, and finish training, oth- ables of the network are the tempering parameters in-
erwise carry out the next step. cluding tempering temperature, tempering time and
5) Calculate all nodes error of the output layer and tempering medium (cooling velocity). As a result, the
hidden layer. number of nodes in the input layer can be 3.
If the jth neuron is in the output layer, δ j is cal- 4.2 Output layer neuron size
culated from As a key elastic component and vibrating sound
δj=yj(1-yj)(yj-y*j) (5) source, the diaphragm of the automotive horn operates in
If the jth neuron is in the hidden layer, δ j is de- an alternating stress state. Furthermore, the failure modes
of the diaphragm involve fatigue fracture and stress re-
termined from
laxation. Therefore, the metal materials used to fabricate
δ j = y j(1 − y j )∑ δk w jk the diaphragm of the automotive horn are expected to
(6)
k possess the excellent mechanical properties, such as high
where δ j is the error of the jth neuron.
elastic limit, high fatigue strength and perfect stress re-
6) Adjust the weights and thresholds according to
laxation resistance. The material parameters which are
the error back-propagation direction.
taken into account in the design of the diaphragm in-
( t +1)
+ ηδ j yi
(t )
wij = wij (7) volves Young’s modulus, elastic limit, yield strength,
tensile strength and elongation. The output variables of
θj ( t +1) =θj ( t ) +ηδj (8)
the network consist of Young’s modulus, elastic limit,
If a momentum coefficient is used, then yield strength, tensile strength and Elongation. Therefore,
wij t +1 = wij t + ηδ j yi + α ( wij t -wij t -1 ) the output layer size of the network is 5.
(9)
4.3 Hidden layer neuron size
t +1 t t t -1
θj = θ j + ηδ j + α[θ j -θ j ] (10) A network with too few hidden nodes is incapable of
where η is the Learning rate, 0 < η < 1 ,and α is differentiating between complex patterns leading to only
the momentum coefficient, 0 < α < 1 . a linear estimate of the actual trend. In contrast, if the
Eq.(7) returns to Eq.(2) to compute repeatedly until network has too many hidden nodes, it will follow the
the error agrees with the prespecified tolerance. noise in the data due to overparameterization leading to
3.4 Normalization of data poor generalization for untrained data. On the other hand,
Normalization of data within a uniform range is with increasing number of hidden nodes, training be-
essential to preventing larger numbers from overriding comes excessively time-consuming. It is shown generally
smaller ones and to preventing premature saturation of that when the number of nodes in the hidden layer is as
hidden nodes, which impedes the learning process. Input two times or so as the number of nodes in the input layer,
and output variables are normalized in interval (0, 1) the network can be more compatible in terms of the ca-
corresponding to the range of the transfer function. pacity and training time. According to the rule, the
However, in order to avoid saturation of the sigmoid number of nodes in the hidden layer is 8.
function leading to slow or no learning between 0 and 0.1 4.4 Learning rate and momentum coeffi-
as well as between 0.9 and 1, it is recommended that the cient
data be normalized between 0.1 and 0.9. The normaliza- A high learning rate will accelerate training by
tion formula is as follows: changing the weight vector. However, this may cause the
Z -Z min (11) search for point with minimum error to oscillate on the
Z n = 0.8 ٛ error surface and never converge, thus increasing the risk
Z max -Z min
794 Vol.24 No.5 JIANG Shuyong et al: Prediction of Mechanical Propertie…