Sie sind auf Seite 1von 3

2009 International Conference on Artificial Intelligence and Computational Intelligence

Optimization of Solidification Process Parameters for


Photosensitive Resin Based on Artificial Intelligence
Yanjue Gong, Fu Zhao, Qiao Bai
College of Mechanical Engineering
Beijing Technology and Business University
Beijing, China
E-mail: gongshangdaxue@hotmail.com

Abstract—This paper presents an optimization approach for photosensitive resin solidified is obtained by using the
photosensitive resin solidification process based on artificial experiment approach. In this paper the rigidity test of SZ-01
neural network combined with orthogonal experiment and UV photosensitive resin is carried out with the HV-1000
genetic algorithm. A predictive model for solidification is microscope rigidity instrument. The test process is that put
established using artificial neural network and the sample for
the poise on the lever of rigidity instrument and make the
neural network model is designed by using orthogonal
experimental method. In the model, the solidification process four pyramid probe press the photosensitive resin sample,
parameters including circumstance temperature, illumination the slim impress appears, and the length of impress is
distance and illumination time are treated as design variables determined with the ruler of microscope, so the rigidity can
and the objective is to obtain the maximum value of rigidity. be calculated depending on the impress length.
Optimization of solidification process parameters for
photosensitive resin was conducted by introducing artificial Ⅲ. METHODS OF INTELLIGENT OPTIMIZATION
neural network prediction models into genetic algorithm. The
results indicate that the optimization method based on
artificial neural network and the genetic algorithm is feasible
A. GENETIC ALGORITHM (GA)
for improve the design quality of the solidification process.
keywords—Optimization, Genetic Algorithm, Artificial A genetic algorithm provides an algorithmic framework
Neural Network , Process Parameters for exploiting heuristics that simulates natural-evolution
processes like selection and mutation. It evolves candidate
I. INTRODUCTION solutions to problems that have large solution spaces and are
Intelligent Optimization which is represented by the not amenable to exhaustive search or traditional
artificial intelligence methods, such as genetic algorithm, optimization techniques. Genetic algorithms have been
artificial neural network has been widely used and focused applied to a broad range of learning and optimization
in the engineering. problems.
With the development of material science, more and more Typically, a genetic algorithm starts with a random
kinds of photosensitive resin appear in the optics component population of encoded candidate solutions, called
manufacture. However the shape and precision of chromosomes. Through recombination process and mutation
photosensitive resin after solidity is easily affected by operators, it evolves the population towards an optimal
solidification process parameters. For a long time, the solution. Generating an optimal solution is not guaranteed
determination of reasonable process parameters depends on and the challenge is thus to design a “genetic” process that
the experience or theory analysis. It may be feasible but not maximizes the likelihood of generating such a solution. The
optimal for the process. So it’s necessary for the resin first step is typically to evaluate the fitness of each
solidification process to optimize the parameters. candidate solution in the current population, and to select
In this paper, a group of optimized process parameters is the fittest candidate solutions to act as parents of the next
found based on intelligent optimization with the methods of generation of candidate solutions. After being selected for
orthogonal experiment, BP Neural Network and Genetic reproduction, parents are recombined (using a crossover
Algorithm optimization, so as to make the rigidity of operator) and mutated (using a mutation operator) to
photosensitive resin arrive its maximum value. generate offspring (see Figure 1 for a description of these
operators). The fittest parents and the new offspring form a
Ⅱ. PHOTOSENSITIVE RESIN RIGIDITY TEST new population, from which the process is repeated to create
new populations.
Because the solidity process of photosensitive resin is
very complex, it’s difficult to determine the formula of
rigidity and process parameters. The rigidity of

978-0-7695-3816-7/09 $26.00 © 2009 IEEE 570


DOI 10.1109/AICI.2009.230
network models, the BP neural model is most widely
applied and can provide good solutions for many industrial
applications, and therefore is used in this paper.
BP neural network is a multi-layer architecture. For the
two layers BP network as shown in Figure 2, the transfer
function of neuron in hidden layer is the sigmoid function
and the transfer function of neuron in output layer is a linear
function. Levenberg-Marquardt rule is used to train the
two-layer BP network. It is developed and trained to fit
Figure 1. Generate offspring functions and make extrapolation. Two learning procedures
Figure 1 is an illustration of crossover and mutation are included in BP network training. The first one is the
operators. Candidate solutions are encoded as strings of bits. positive propagation process in which input signal is
Parents a and b are recombined to produce offspring c and d, transferred layer by layer and practical outputs of every
a crossover is performed at the 4th bit, i.e., the tails of both neurons are computed; the second procedure is the back
parents are swapped starting from the 4th bit. Offspring c is propagation in which the errors between practical and
then mutated to produce c’: a mutation is performed at the expected outputs are progressively computed layer by layer,
3rd bit, i.e., the value of the 3rd bit is flipped[1]. and weights are adjusted according to the errors[2-3].
To illustrate an iteration of a genetic algorithm, consider
the Boolean satisfiability problem. Assume that we want to
find a satisfying assignment to the following Boolean
formula: (x1 ∨x2 ∨¬x3)∧(x2 ∨x3 ∨x4)∧(¬x2 ∨x5 ∨¬x6)∧
(¬x4∨¬x5∨¬x6)∧(x3∨¬x5∨¬x6)∧(x3∨x4∨x5). Let’ s say
Figure 2. Two layers BP neural network
we have two (randomly generated) candidate solutions, a :
{x1 = 1, x2 = 1, x3 = 0, x4 = 0, x5 = 0, x6 = 1} and b : {x1 = 0,
x2 = 0, x3 = 0, x4 = 0, x5 = 1, x6 = 1}. If we evaluate the Ⅳ. OPTIMUM DESIGN
formula on a, we see that clauses 3 and 6 are false, whereas
evaluating the formula on b makes clauses 2 and 5 false. In order to find the rational combination of process
Both a and b are not satisfying assignments. We now parameters which make the rigidity of resin reach maximum
recombine a and b to produce an offspring c : {x1 = 1, x2 = 1, value, the optimum design process is followed: The rigidity
x3 =0, x4 = 0, x5 = 1, x6 = 1}, which takes the first three data are obtained by the method of orthogonal experiment.
variable assignments from a and the last three from b. Based on these data, the relationship of the rigidity and
Offspring c does not define a satisfying assignment either process parameters is found by BP neural network.
since it makes clause 5 false. However, if we mutate the According to this relationship, the rational combination of
value c assigns to x3 to produce d: {x1 = 1, x2 = 1, x3 = 1, x4 process parameters is calculated by using of Genetic
= 0, x5 = 1, x6 = 1}, we see that d does provide a satisfying Algorithm optimization.
assignment to our Boolean formula.
The operations of evaluation, selection, recombination A. Orthogonal Experiment Method
and mutation are usually performed many times in a genetic
algorithm. Selection, recombination, and mutation are There are three process parameters in photosensitive resin
generic operations in any genetic algorithm and have been solidity process including circumstance temperature,
thoroughly investigated in the literature. On the other hand, illumination distance and illumination time. Since every
evaluation is problem specific and relates directly to the technology parameter has different value, it will spend
structure of the solutions (i.e., how candidate solutions are redundant time making a lot of experiments with different
encoded as chromosomes and related to each other). combinations of these parameters. Therefore the timesaving
Therefore, in a genetic algorithm, a major issue is the choice and efficient orthogonal experiment method is applied.
of the solutions structure and the evaluation (fitness function) The standard orthogonal experiment table L9(34) is
method. Other parameters include the size of the population, choose, in which there are four parameters, and three level
the portion of the population taking part in recombination, in every parameter, totally ninth times experiments. In this
and the mutation rate. The mutation rate defines the paper there are only three process parameters that are
probability with which a bit is flipped in a chromosome that circumstance temperature T, illumination distance s and
is produced by a crossover [1]. illumination time t. Each parameter and level is described as
table 1.
TABLE Ⅰ. PARAMETERS AND LEVELS
B. BP NEURAL NETWORK
Levels T/ ℃ s / mm t / min
Level 1 20 50 120
Neural networks mimic the way by which biological Level 2 40 100 220
brain neurons generate intelligent decisions. Among neural Level 3 60 200 320

571
With the HV-1000 microscope rigidity instrument the temperature x1/ ℃ , illumination distance x2/cm and
rigidity of different process parameters combinations is illumination time x3/min, and the boundary condition of
obtained as shown in table 2. every parameter is defined as below:
20 ≤ x1 ≤ 60
TABLE Ⅱ. EXPERIMENT RESULTS 50 ≤ x 2 ≤ 200
NO. Level 1 Level 2 Level 3 Rigidity
1 1 1 1 26.89 120 ≤ x3 ≤ 320
2 1 2 2 12.69
3 1 3 3 10.71 D. Optimum Parameters
4 2 1 2 30.41
5 2 2 3 25.48
6 2 3 1 16.5 The number of seed is set as 80, the length of Genetic
7 3 1 3 20.64 code is 20, the ratio of variation is 0.01 and the ration of
8 3 2 1 31.48 cross is 0.95 in Genetic Algorithm optimization.
9 3 3 2 18.60

E. The Results
B. Optimum Design Object Function
With the calculation of MATLAB[4-6], the results are
The mathematics formula about rigidity and process
shown as table 3.
parameters is difficult to set up, but the relationship of TABLE Ⅲ. OPTIMUM DESIGN RESULTS
them can be found with BP neural network. The chosen BP Third experiment Optimum
neural network has three layers, and there are three neural Parameters
parameters parameters
units in input layer, ten neural units in middle layer and one x1/℃ 60 60
neural unit in output layer. The data of table 2 are seen as x2/mm 125 91
training sample, with the calculation of MATLAB, the x3/min 120 245
Rigidity 31.48 38.2
relationship of rigidity and process parameters is get after 45
times training, the convergence process is shown as figure 3.
From the optimal design results, the maximum rigidity is
38.2 after optimum design, and the rational combination of
process parameters is determined that make the rigidity of
photosensitive resin reach maximum value with Genetic
Algorithm optimization.

Ⅴ. CONCLUSION

With the methods of orthogonal experiment and BP


Neural Network, the relationship of the rigidity and process
parameters is set up. Using this rule a group of reasonable
process parameters is found by Genetic Algorithm
optimization, which can make the rigidity of photosensitive
resin reach the maximum value.

Figure 3. Training process Reference


[1]Patrice godefroid, Sarfraz Khurshid. “Exploring very large state spaces
using genetic algorithms”. Proceeding of TACAS, pp. 266-280, 2002.
The object function is the relationship obtained from BP [2]X-X. Li, H.Huang, C-H. Liu. “The application of an ANFIS and BP
neural network, and the optimized design aim is to find the neural networks method in vehicle shift decision”. Proceeding of 12th
maximum value of rigidity. There are a lot of iterative IFToMM World Congress, Besancon(France), pp. 25-29,2007.
processes in optimum design, and if BP neural network is [3]Chih-Ming Hsu. “Solving multi-response problems through neural
networks and principal component analysis”. Journal of the Chinese
trained in every iterative process, it would cost a lot of time Institute of Industrial Engineers, Vol. 18,No. 5,pp. 47-54.
and make the BP neural network over training. So the only [4]Zhao xuanmin. Method of Experiment Design. Beijing, Science
BP neural network weight matrix is adopted in every Publishing(2006).
iterative process, it performs with very high efficiency. [5]Simon Haykin. Neural Network Theory. Beijing, Mechanical
Engineering Publishing(2004).
[6] Zhang wei. Basic and Program of Matlab. Xi’an, Xi’an electronic
C. Design Parameters and Boundary Conditions science university Publishing(2004).

There are three design parameters that are circumstance

572

Das könnte Ihnen auch gefallen