You are on page 1of 9

Journal of Materials Processing Technology 170 (2005) 602610

Multi-response optimization in friction properties of PBT composites using Taguchi method and principle component analysis
Chin-Ping Fung , Po-Chung Kang
Department of System Engineering, Chung Cheng Institute of Technology, Tao-Yuan 33509, Taiwan, ROC Received 27 November 2003; received in revised form 20 February 2004; accepted 1 June 2005

Abstract This paper optimizes the injection-molding process for friction properties of ber-reinforced polybutylene terephthalate using Taguchi method and principal component analysis. Four controllable factors of the manufacturing process were studied at three levels each. The single-response optimization of friction property was conducted by Taguchi method. Principal component analysis was employed to correspond to multi-response cases, for transforming the correlated friction properties to a set of uncorrelated components and evaluating the principal components. The optimum combination of process factors and levels for multiple qualities, based on the rst principal component, was determined rst. The appropriate number of the principle components, and the inuence of the number on the optimum process condition, were subsequently studied by extracting more than one principal component and integrating into a comprehensive index. Finally, the analysis of variance was used to nd out the most inuential injection-molding parameter for single and multiple responses problems. 2005 Elsevier B.V. All rights reserved.
Keywords: PBT; Friction properties; Injection molding; Multi-response optimization; Taguchi method; Principal component analysis

1. Introduction Reinforced plastics have attracted much interest in structural applications for many years. Among these plastics, ber-reinforced polybutylene terephthalate (PBT) is a very frequently used material. PBT molding composites are designed for engineered parts that require heat resistance, dimensional stability, rigidity and strength, chemical resistance and exceptional toughness. Typical applications include electrical connectors, switches, automotive underhood components, electrical appliance components and outdoor telecommunications enclosures. Investigating PBT composites has thus received some attention. Researchers in this area have been particularly interested in tensile strength [1], fracture performance [2], fatigue crack propagation [3], and adhesive wear [4]. The literature published to date has examined the effects on the mechanical performance of PBT

Corresponding author. Tel.: +886 3 3809257; fax: +886 3 3906385. E-mail address: cpfung@ccit.edu.tw (C.-P. Fung).

composites of material parameters, such as the ber length, volume fraction, and ber concentration. However, neither the optimization of injection-molding process nor the inuence of process parameters on the mechanical properties of PBT has yet been studied. Injection molding is a very common manufacturing process for reinforced plastics because it provides products with dimensions steadiness and precision, a low manufacture cycle and low cost. However, the author of this study [5,6] and other researchers [79] have found that some parameters of injection-molding process greatly affect the quality of products. Selecting the correct injectionmolding conditions is always a major concern to the plastics industry. Traditional experiment design would require many experimental runs to achieve satisfactory result. In answer to this problem, Taguchi [10] devised a new experiment design that applied signal-to-noise ratio and orthogonal arrays to the robust design of products and process. In recent years, Taguchi method has become a widely accepted methodology for improving productivity. However, most research based on Taguchi method has been concerned with optimizing

0924-0136/$ see front matter 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.jmatprotec.2005.06.040

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610

603

only a single response or quality characteristic. It is not the kind of condition that engineers face in todays complex manufacturing processes. To optimize several responses or quality characteristics simultaneously, many researchers have tried to combine Taguchi method with other methods [1114]. Logothetis and Haigh [11] employed a regression technique to optimize the combination of process factors and levels. This approach actually increases the complexity of computational process, without considering correlations among the responses. Reddy et al. [12] proposed another method for optimizing three responses in a certain injection-molding process. Within Taguchis methodology, this was a goal-programming approach. However, any engineer with limited statistical knowledge could nd this method too difcult to understand and therefore, it is not widely accepted. The methodology [13,14] that utilizes both Taguchi method and principal component analysis (PCA) is a quite practical and effective procedure for tackling multi-response problems. By using PCA, a set of original responses is transformed into a set of uncorrelated components so that the optimal factor/level combination can be found. However, those cases studied in the published literature has had one eigenvalue larger than one; and thus, the rst principal component has been the only component extracted. This does not occur in the majority of cases in todays complex manufacturing processes we have. In particular, when more than one principle component is considered, it is doubtful whether the factor/level combination, determined by the rst principal component only, is optimal or not. Therefore, the optimizing procedure must correspond to cases that have more than one eigenvalue larger than one and more than one principal component extracted. This paper, using Taguchi method and PCA, presents the optimum levels of injection-molding parameters for the friction properties of ber-reinforced polybutylene terephthalate. The experimental design used orthogonal array L9 for the four controllable factors of lling time, melt temperature, mold temperature and packing pressure, each at three levels to nd the optimum combination of factors and levels in manufacturing process. The single-response optimization of friction property was conducted using Taguchi method. For a multi-response case, PCA was employed to transform the correlated friction properties to a set of uncorrelated components and then to evaluate the principal components. Firstly, the optimum process factor/level combination based on the rst principal component was determined. In addition, since the case in this study has more than one eigenvalue larger than one, the appropriate number of the principle components and the impact of that number on the optimum process condition, were thus studied by extracting more than one principal component and integrating them into a comprehensive index. Finally, the analysis of variance (ANOVA) was used to nd out the most inuential injectionmolding parameter for problems of single and multiple responses.

2. Analysis methods 2.1. Taguchi method and signal-to-noise ratio Taguchi method is a widely accepted methodology for contemporary experiment design. In this method, signalto-noise (S/N) is used to represent a response or quality characteristic and the largest S/N ratio is required. There are usually three types of quality characteristics, i.e. target-the-best, larger-the-better and smaller-thebetter. 1. target-the-best S = 10 log N 2. larger-the-better S = 10 log N 1 n 1 2 y i=1 i
n i=1 2 yi n

1 nS

n i=1 2 yi

= 10 log

y 2 S2

(1)

(2)

3. smaller-the-better S = 10 log N 1 n = 10 log( y2 ) (3)

where S is the standard deviation, yi is the measured property, and n corresponds to the number of samples in each test trial. 2.2. Principal component analysis The principal component analysis (PCA) is a multivariate statistical method that selects a small number of components to account for the variance of original multi-response. The technique was rst introduced by Pearson [15] and then further developed by Hotelling [16]. The procedure is described as follows: 1. The original multi-response array xi (j ), i = 1, 2, . . . , m; j = 1, 2, . . . , n x1 (1) x1 (2) x2 (2) . . . . . . xm (2) x1 (n) x2 (n) . . . . . .

x (1) 2 . . X= . . . . xm (1)

(4)

xm (n)

where m is the number of test trial and n is the number of the response. In this paper, X is the S/N ratio of each response. 2. Normalizing the response

604

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610

The response is normalized using the following formula to get rid of the difference between units.
(j ) xi

3. Experimental design and results 3.1. Material A commercial 30 wt.% short glass ber-reinforced polybutylene terephthalate (PBT) Valox 310SEO from General Electric, USA, was used. Table 1 gives its basic physical properties. In order to remove moisture, the composite was heated to 120 C for 4 h prior to the process of injection molding.

xi (j ) xi (j ) = xi (j )+ xi (j )
(1) x1 (2) x1

x (1) x (2) x (n) 2 2 2 . . . . . . . . X = . . . . . . . . . . (1) x (2) x (n) xm m m

(n) x1

(5) 3.2. Experimental design The experiment was conducted with four controllable three-level factors of manufacturing process and four responses. Four factors were the lling time, melt temperature, mold temperature, and ram speed. The injection molding conditions are listed in Table 2. The four responses, or quality characteristics, were the friction coefcients and surface roughness in different sliding directions. Normally, the fullfactorial design would require 34 = 81 experimental runs. However, the effort and experimental cost required for such a design could be prohibitive and unrealistic. In this study, nine experimental runs based on the orthogonal array L9 (34 ) as shown in Table 3, were conducted to study the four process factors. The specimens were molded in a computerized reciprocating screw injection-molding machine. The dimensions of the mold cavities as well as the location and dimensions of the pin gate were simulated using C-MOLD rst to achieve a better lling process. 3.3. Friction test Friction tests were carried out with a Schwingum Reibung Verschleiss (SRV) oscillation friction wear tester. Fig. 1 shows a diagram of the schematic mechanism of the tester. The specimens were 15 mm 15 mm 4.0 mm, cut from the center part of the injection-molding specimens. The tests were performed under ball-on-plane contact with a load of 60 N. The ball was a chromium steel ball (AISI E52100), 10 mm in diameter with an average hardness of 62 2 HRC. The stroke was 2 mm, and the frequency and test duration were 25 Hz and 30 min, respectively. The above stroke, frequency, and test duration resulted in a total sliding distance of
Table 1 Basic physical Properties of PBT (Valox 310SEO) Properties Density Mold shrinkage (%) Tensile strength (MPa) Flextural modulus (MPa) Tensile strain (%) Izod impact strength (kJ/m2 ) Coefcient of linear thermal expansion (1/ C) Thermal conductivity (W/m C) Vicat softening temperature ( C) (g/cm3 ) Valox 310SEO 1.41 1.11.8 55 2600 20 5 10 105 0.24 212

(j ) is the normalized response, x (j)+ is the maxwhere xi i imum of xi (j), and xi (j) is the minimum of xi (j). 3. Correlation coefcient array The correlation coefcient array of the normalized response array is evaluated as follows:

Rjl =

(j ), x (l)) Cov(xi i xi (j) xi (l)

, (6)

j = 1, 2, . . . , n; l = 1, 2, . . . , n

(j ), x (l)): The covariance of sequences where Cov(xi i xi (j ) and xi (l); xi (l) : The standard deviation of sequence (l). xi 4. Determining the eigenvalues and eigenvectors The eigenvalues and eigenvectors are determined from the correlation coefcient array,

(R k Im )Vik = 0
n

(7) k = n,
k=1

where k : eigenvalues,

k = 1, 2 , . . . , n ;

Vik = [ak1 ak2 . . . akn ]T : eigenvectors corresponding to the eigenvalue k . 5. Evaluating the principal components We have the following uncorrelated principal components:
n

Ymk =
i=1

Xm (i) Vik

(8)

The principal components are created in order of decreasing variance, and therefore the rst principal component, Ym1 , accounts for most variance in the data. Based on Kaisers study [17], the components with an eigenvalue greater than one are chosen to replace the original responses for further analysis. 6. Evaluating the coefcient of determination Ck = k , n k = 1, 2, . . . , n (9)

The coefcient of determination, Ck , represents the weight of the principle component, Ymk .

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610 Table 2 Experimental factors and levels for the injection mold conditions of PBT Levels of experimental factor Experimental factors A Filling time (s) 1 2 3 2 3 7 B Melt temperature ( C) 240 260 280 C Mold temperature ( C) 60 90 120

605

D Ram speed (%) 50 75 100

Table 3 Orthogonal array L9 (34 ) of the experimental runs for 310SEO Experimental run 1 2 3 4 5 6 7 8 9 A Filling time (s) 1 1 1 2 2 2 3 3 3 B Melt temperature ( C) 1 2 3 1 2 3 1 2 3 C Mold temperature ( C) 1 2 3 2 3 1 3 1 2 D Ram speed (%) 1 2 3 3 1 2 3 2 1

180 m. The sliding direction of the SRV wear test is dened as either parallel or perpendicular to the injection melt ow. The former is called the P-type, the latter the AP-type. 3.4. Friction coefcient and surface roughness The friction coefcient was obtained through the stable part of the variation of friction coefcient that was exported from SRV wear tester. The surface roughness was measured using a surface analyzer and represented as the roughness average (Ra ). Table 4 lists all the experimental results. It can be seen that the results of some experimental runs were very good in one sliding direction, but that the opposite results happened in the other sliding direction. It is important, therefore, to optimize the process factors of injection molding for good friction properties in both sliding directions.

Table 4 The experimental results of friction coefcient () and surface roughness average (Ra ) under different sliding directions for 310SEO Experimental run 1 2 3 4 5 6 7 8 9 Friction coefcient () P-type 0.24 0.024 0.26 0.023 0.20 0.014 0.24 0.023 0.26 0.021 0.21 0.023 0.25 0.026 0.24 0.028 0.25 0.025 AP-type 0.26 0.032 0.25 0.025 0.24 0.032 0.25 0.045 0.25 0.030 0.23 0.029 0.25 0.028 0.22 0.029 0.26 0.030 Surface roughness (Ra ) P-type 0.57 0.77 0.49 2.16 0.56 0.76 0.60 0.52 0.56 AP-type 1.03 1.21 0.40 2.23 0.26 1.52 1.10 1.69 0.24

4. Analysis and discussion The response of friction properties of ber-reinforced PBT composites is highly dependent on the conditions of the injection-molding process. During the process, the main design factors such as lling time, melt temperature, mold temperature, and ram speed can greatly affect the ber condition and change the mechanical properties. In the previous study [5], the inuence of various injection molding process factors on the tribological properties of ber-reinforced PBT was investigated. The study showed that the friction properties are highly dependent on the solidied-intermediate-core morphology that is dened by ber orientation and exists from free surface to core. The friction coefcient is lowest when the thickness of solidied layer is smallest. The solidied layer was a result of the interaction between the melt ow and mold wall. In the process of injection-molding, the thickness of solidied layer is determined by the main design

Fig. 1. The schematic mechanism of SRV oscillation friction wear tester.

606

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610

factors as these factors determine the status of melt ow such as velocity gradient, cooling effect, injection pressure and heat dissipation. The process of optimization in this study was to determine the levels of design variables that reduce the variation of product quality from their target value. It was a statistical process used to keep manufacturing under control. According to Taguchis methodology, no matter how the quality of the product is measured, the quality characteristics are divided into three characteristics: target-the-best, larger-the-better and smaller-the-better. When measures of quality for example, friction coefcients and surface roughness in this study become worse as the value of the measure increases, measures with this property are said to possess smaller-the-better characteristics. Based on the smaller-the-better methodology, the problems of single-response optimization and multi-response optimization were thus analyzed as follows. 4.1. Single-response problem Four friction properties, i.e. friction coefcients and surface roughness in different sliding directions (P-type and AP-type), were separately selected to be the target quality in this study. When a single-response problem is considered, Taguchi method can be employed to obtain the optimal level/factor combination of injection-molding process. The signal-to-noise is used to represent quality characteristic and the largest S/N ratio is demanded. In addition, the contribution of each factor to the friction quality can be evaluated using the analysis of variance (ANOVA). A smaller friction coefcient or surface roughness is normally required in most industry cases. Therefore, the smallerthe-better methodology of S/N ratio was employed for the optimization of friction property. The S/N ratios of four friction properties of the nine experimental runs are listed in Table 5. The response table of S/N ratios of the P-type friction coefcient was calculated, as shown in Table 6. The S/N ratio of factors AD is maximum at A1 , B3 , C1 and D3 , respectively. As a result, the factor/level combination A1 B3 C1 D3 was recommended. The results of ANOVA for the S/N ratios of the P-type friction coefcient are also shown in Table 6. It can be seen that the contribution of factor B to the P-type
Table 5 The S/N ratios of experimental results Experimental run Friction coefcient () P-type 1 2 3 4 5 6 7 8 9 12.40 11.70 13.98 12.40 11.70 13.56 12.04 12.40 12.04 AP-type 11.70 12.04 12.40 12.04 12.04 12.77 12.04 13.15 11.70 Surface roughness (Ra ) P-type 4.88 2.27 6.20 6.69 5.04 2.38 4.44 5.68 5.04 AP-type 0.26 1.66 7.96 6.97 11.70 3.64 0.83 4.56 12.40

friction coefcient was the largest (41.5%). The melt temperature (factor B) was the most important factor to the P-type friction coefcient. The same analysis procedure can also be applied to optimize the injection-molding conditions for the AP-type friction coefcient, P-type surface roughness, and AP-type surface roughness, respectively. Their response tables of their S/N ratios are listed in Tables 79, respectively. The levels that gave the largest average response were selected from the response tables. They were the level/factor combinations of A3 B2 C1 D2 for the AP-type friction coefcient, A3 B3 C3 D1 for the P-type surface roughness, and A3 B3 C3 D1 for the APtype surface roughness. The results of ANOVA in Tables 79 show the ram speed was the most important factor for both the AP-type friction coefcient and surface roughness. However, the mold temperature was the most important factor for the P-type surface roughness. It can also be seen that the lling time was quite an important factor to the P-type surface roughness, but it was a negligible factor in other qualities because of its contribution. It was therefore found that the optimal factor/level combination, or the most important factor, for one quality was usually different from that for another quality. In such a case, an engineering judgment that refers to past experience is the only real guarantee of correct decisionmaking in the injection-molding process. 4.2. Multi-response problem When the single-response problem is considered, as discussed in the previous section, one factor could be very important to one quality characteristic while it may be unimportant to the other quality characteristics in the process of injection molding. However, most cases in industrial products present multi-response problems. The many different properties of a product are always simultaneously considered for multi-functional purposes. It is necessary to optimize the injection-molding conditions for multi-response problems. The PCA and Taguchi method were integrated in this study to deal with multi-response problems. First, the multiresponse array, in which the elements were the S/N ratios of each response, was normalized using Eq. (5). The results are listed in Table 10. Then, the correlation coefcient matrix was evaluated from the normalized response array. Table 11 lists the eigenvalues, which were determined using Eq. (7). There are two of four eigenvalues larger than one. The eigenvector that corresponded to the largest eigenvalue 1.718 was [0.62, 0.87, 0.56, 1.00]. The other eigenvectors were [0.79, 0.66, 1.00, 0.51], [1.00, 0.64, 0.57, 0.38], and [0.32, 0.85, 0.82, 1.00], that corresponded to the eigenvalues 1.437, 0.626, and 0.220, respectively. The four elements of the eigenvector are the weights of the four responses. The rst principal component was the sum of the products of the S/N ratios of four responses multiplied by the elements of the eigenvector, which corresponded to the largest eigenvalue. The other principal components were obtained in the same way. The results are all listed in Table 10.

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610 Table 6 The response table and ANOVA of S/N ratios of P-type friction coefcients Factor Levels 1 A B C D Total Error Note: Values enclosed in parentheses represent standard deviation. Table 7 The response table and ANOVA of S/N ratios of AP-type friction coefcients Factor Levels 1 A B C D Total Error 12.046 (0.348) 11.928 (0.197) 12.539 (0.752) 11.814 (0.197) 2 12.283 (0.418) 12.411 (0.641) 11.928 (0.197) 12.653 (0.564) 3 12.298 (0.759) 12.287 (0.541) 12.159 (0.205) 12.159 (0.205) 2 2 2 2 8 2 0.120 0.379 0.572 1.066 1.863 0.119 0.060 0.189 0.286 0.533 0.060 1.000 3.162 4.776 8.901 0.0 13.9 24.3 50.8 100 Degree of freedom Sum of square Mean square F value 12.692 (1.168) 12.278 (0.205) 12.782 (0.670) 12.046 (0.348) 2 12.551 (0.937) 11.932 (0.401) 12.046 (0.348) 12.551 (0.937) 3 12.159 (0.205) 13.192 (1.019) 12.574 (1.229) 12.805 (1.032) 2 2 2 2 8 2 0.457 2.543 0.865 0.897 5.025 0.457 0.228 1.271 0.432 0.448 0.228 1.001 5.576 1.896 1.967 0.01 41.5 8.1 8.8 100 Degree of freedom Sum of square Mean square F value

607

Contribution (%)

Contribution (%)

In previous published studies [13,14], only the rst principal component has been chosen to represent the original responses since one eigenvalue has been larger than one. However, this is no longer relevant to the majority of cases in todays complex manufacturing processes. In this study, two of four eigenvalues were greater than one. Thus, comparing different optimal process factor/level combinations, as determined by extracting different number of principal components, is a major concern. To integrate more than one principal component to a comprehensive index, a simple
Table 8 The response table and ANOVA of S/N ratios of P-type surface average roughness Factor Levels 1 A B C D Total Error 4.450 (1.998) 0.877 (6.556) 4.315 (1.720) 4.985 (0.089) 2 0.244 (6.149) 4.329 (1.812) 0.206 (6.129) 3.445 (1.937) 3 5.051 (0.622) 4.539 (1.954) 5.223 (0.894) 1.315 (6.987) 2 2 2 2 8 2

approach based on the coefcient of determination was used. Since the coefcient of determination represents the weight of the principle component, the comprehensive index was obtained from the sum of the products of the principle component multiplied by the coefcient of determination. The results are listed in Table 12. First, only the rst principal component was considered. The response table and ANOVA for the rst principal component of four quality characteristics are listed in Table 13. It can be seen that when multiple quality char-

Degree of freedom

Sum of square

Mean square

F value

Contribution (%)

41.163 25.370 42.886 20.381 125.535 20.381

20.582 12.685 21.443 10.190 10.190

2.020 1.245 2.104 1.000

16.6 4.0 17.9 0.00 100

Table 9 The response table and ANOVA of S/N ratios of AP-type surface average roughness Factor Levels 1 A B C D Total Error 2.015 (5.194) 2.684 (3.719) 2.817 (2.265) 7.947 (7.113) 2 0.366 (9.956) 1.829 (8.671) 1.258 (10.00) 3.283 (1.483) 3 2.337 (8.909) 5.573 (8.278) 6.277 (6.431) 0.055 (7.501) 2 2 2 2 8 2 6.709 102.541 124.504 199.534 417.659 6.709 3.354 51.271 62.252 99.767 3.354 1.000 15.285 18.559 29.742 0.0 22.9 28.2 46.2 100 Degree of freedom Sum of square Mean square F value Contribution (%)

608

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610

Table 10 The principal components of four quality responses Experimental run Normalized multi-response array (S/N) (P) 1 2 3 4 5 6 7 8 9 0.693 1.000 0.000 0.693 1.000 0.184 0.851 0.693 0.851 (AP) 1.000 0.766 0.517 0.766 0.766 0.262 0.766 0.000 1.000 Ra (P) 0.102 0.305 0.000 1.000 0.090 0.296 0.137 0.040 0.090 Ra (AP) 0.654 0.726 0.229 1.000 0.036 0.828 0.683 0.876 0.000 Principal component First 0.589 0.389 0.221 0.464 1.199 0.652 0.434 0.469 1.347 Second 1.643 1.970 0.458 2.563 1.404 1.037 1.662 1.034 1.422 Third 0.243 0.612 0.244 0.013 0.473 0.162 0.543 1.003 0.160 Fourth 1.198 0.807 0.669 0.609 0.293 0.749 0.949 0.621 0.504

Table 11 Eigenvalues for the principal components Principal component First Second Third Fourth Eigenvalue 1.718 1.437 0.626 0.220

acteristics were considered, A2 B1 C1 D3 was recommended and ram speed was the most important factor. The contribution from ram speed was 64.7%. In addition, the recommended process factor/level combination was different from the combinations, proposed in the previous section, when the single-response problem is considered. This plainly indicates the difference between single-response and multi-response problems. Next, the rst and the second principle components that correspond to the two eigenvalues larger than one, were

considered. The response table and ANOVA are listed in Table 14. It can be seen that A2 B1 C2 D3 was recommended. The recommended level of mold temperature was C2 . It was different from the previous recommended condition in which only the rst principal component was considered. This indicates that besides the rst principal component corresponding to the largest eigenvalue, the other principal components corresponding to eigenvalues larger than one are also crucial for obtaining the optimal process condition. Ram speed was still the most important factor. The contribution from ram speed was 25.5%. On the other hand, when all the principle components were considered, the combination A2 B1 C2 D3 was recommended, as shown in Table 15. It is the same condition as that from which the rst two principle components, corresponding to the two eigenvalues larger than one, were extracted. It indicates that the optimal condition is not affected by the two principle components that corresponded to the eigenvalues smaller than one.

Table 12 The comprehensive index that obtained by extracting different number of principal components Experimental run Comprehensive index First principal component 1 2 3 4 5 6 7 8 9 0.589 0.389 0.221 0.464 1.199 0.652 0.434 0.469 1.347 First + second principal components 0.337 0.540 0.070 1.120 0.011 0.652 0.411 0.573 0.068 Fourth principal components 0.441 0.681 0.068 1.155 0.079 0.719 0.548 0.764 0.015

Table 13 The response table and ANOVA for the rst principal component Factor Levels 1 A B C D Total Error 0.400 (0.184) 0.186 (0.569) 0.177 (0.670) 1.045 (0.402) 2 0.028 (1.019) 0.373 (0.834) 0.424 (0.906) 0.057 (0.614) 3 0.438 (0.908) 0.305 (1.002) 0.618 (0.514) 0.237 (0.397) 2 2 2 2 8 2 0.308 0.054 1.032 2.707 4.101 0.054 0.154 0.027 0.516 1.354 0.027 5.705 19.130 50.202 6.187 0.0 23.844 64.708 100 Degree of freedom Sum of square Mean square F value Contribution (%)

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610 Table 14 The response table and ANOVA for the rst two principal components Factor Levels 1 A B C D Total Error Table 15 The response table and ANOVA for four principal components Factor Levels 1 A B C D Total Error 0.397 (0.308) 0.715 (0.385) 0.641 (0.175) 0.168 (0.241) 2 0.651 (0.541) 0.508 (0.374) 0.607 (0.589) 0.649 (0.090) 3 0.432 (0.402) 0.257 (0.402) 0.232 (0.274) 0.662 (0.550) 2 2 2 2 8 2 0.114 0.315 0.310 0.475 1.214 0.114 0.057 0.157 0.155 0.238 0.057 2.762 2.718 4.172 16.538 16.131 29.779 100 Degree of freedom Sum of square Mean square F value 0.316 (0.236) 0.623 (0.432) 0.521 (0.164) 0.086 (0.219) 2 0.587 (0.568) 0.367 (0.328) 0.531 (0.594) 0.534 (0.121) 3 0.305 (0.333) 0.218 (0.382) 0.156 (0.224) 0.587 (0.525) 2 2 2 2 8 2 0.153 0.251 0.273 0.455 1.132 0.153 0.077 0.125 0.136 0.227 0.077 1.639 1.782 2.971 8.639 10.585 26.662 100 Degree of freedom Sum of square Mean square F value

609

Contribution (%)

Contribution (%)

Table 16 Prediction of S/N ratios of process conditions extracting different number of principal components Quality characteristic S/N ratios of process conditions First principal component Friction coefcient (P-type) Friction coefcient (AP-type) Surface roughness (P-type) Surface roughness (AP-type) 1.7884 1.7526 0.4867 1.4773 First + second principal components 0.4582 0.3804 1.6413 0.6507 Fourth principal components 0.6466 0.6874 1.9483 0.9577

To compare the anticipated improvement between the different recommended process combinations, the signal-tonoise ratios of four qualities were predicted using the additive model. Table 16 lists the comparison of results. It can be seen that the process conditions that extracted more than one principal component had larger S/N ratios. It indicates the process condition that extracts only the rst principal component is not the optimal condition. Consequently, the procedure of process optimization, if it is to be thoroughly considered, should extract at least all the principal components that correspond to the eigenvalues larger than one.

5. Conclusions Taguchi method and the principle component analysis were used in this study to optimize the injection-molding conditions of ber-reinforced PBT. The results are summarized as follows: 1. Various optimum injection-molding conditions were found for the friction properties in a single-response problem. The factor/level combination A1 B3 C1 D3 , A3 B2 C1 D2 , A3 B3 C3 D1 , and A3 B3 C3 D1 were recom-

mended for the P-type friction coefcient, AP-type friction coefcient, P-type surface roughness, and AP-type surface roughness, respectively. 2. In the multi-response problem, friction coefcients and surface roughness in different sliding directions (Ptype and AP-type) were simultaneously considered. A2 B1 C1 D3 was recommended for the condition in which only the rst principal component is extracted. However, A2 B1 C2 D3 was recommended for the conditions in which more than one principal component is extracted. 3. The anticipated improvement for different recommended process conditions showed that those process conditions extracting more than one principal components have larger S/N ratios. The process condition that extracts only the rst principal component is thus not the optimal condition. Consequently, the procedure of process optimization should extract at least all the principal components that correspond to the eigenvalues that are larger than one.

Acknowledgements The authors would like to acknowledge the assistance of Tse-Chang Li for his help with friction test. The authors

610

C.-P. Fung, P.-C. Kang / Journal of Materials Processing Technology 170 (2005) 602610 ites in iniection molding, J. Reinf. Plast. Composite 19 (2000) 301321. K.C. Ho, M.C. Jeng, Fibre orientation of short glass bre reinforced polycarbonate composites under various injection moulding conditions, Plast. Rubb. Composite Proc. Appl. 25 (1996) 469476. K.C. Ho, J.R. Hwang, J.L. Doong, Tensile properties of short glass bre reinforced polycarbonate, Polym. Polym. Composite 4 (1996) 563575. G. Taguchi, Introduction to Quality Engineering, Asian Productivity Organization, Tokyo, 1990. N. Logothetis, A. Haigh, Characterizing and optimizing multiresponse processes by the Taguchi method, Qual. Reliab. Eng. Int. 4 (1988) 159169. P.B.S. Reddy, K. Nishina, A. Subash Babu, Unifcation of robust design and goal programming for multiresponse optimisationa case study, Qual. Reliab. Eng. Int. 13 (1997) 371383. C.T. Su, L.I. Tong, Multi-response robust design by principal component analysis, Total Qual. Manage. 8 (1997) 409416. J. Antony, Multi-response optimization in industrial experiments using Taguchis quality loss function and principal component analysis, Qual. Reliab. Eng. Int. 16 (2000) 38. K. Pearson, On lines and planes of closest t to systems of points in spaces, Philos. Mag. Series 6 2 (1901) 559572. H. Hotelling, Analysis of a complex of statistical variables into principal components, J. Educ. Psychol. 24 (1933) 417441. H.F. Kaiser, The application of electronic computers to factor analysis, Educ. Psychol. Meas. 20 (1960) 141151.

would also like to acknowledge the help of Cheng-Hung Huang of the United Ship Design and Development Center, Taiwan, for his helpful discussions. References
[1] J. Denault, T. Vu-Khanh, B. Foster, Tensile properties of injection molded long ber thermoplastic composites, Polym. Composite 10 (1989) 313321. [2] T. Vu-Khanh, J. Denault, Fracture behavior of long ber reinforced thermoplastics, J. Mater. Sci. 29 (1994) 57325738. [3] H. Voss, J. Karger-Kocsis, Fatigue crack propagation in glass-bre and glass-sphere lled PBT composites, Int. J. Fatigue 10 (1988) 311. [4] C. Lhymn, Tribological properties of poly-(butylene terephthalate) glass composites: adhesive wear, Mater. Sci. Eng. 80 (1986) 93100. [5] M.C. Jeng, C.P. Fung, T.C. Li, The study on the tribological properties of ber-reinforced PBT composites for various injection molding process parameters, Wear 252 (2002) 934945. [6] C.P. Fung, J.R. Hwang, C.C. Hsu, The effect of injection molding process parameters on the tensile properties of short glass berreinforced PBT, Polym. Plast. Technol. Eng. 42 (2003) 4563. [7] S.H. Chang, J.R. Hwang, J.L. Doong, Manufacturing process optimization of short glass ber reinforced polycarbonate compos-

[8]

[9]

[10] [11]

[12]

[13] [14]

[15] [16] [17]

You might also like