The MotoTech Manufacturing Company: Design of Experiments
Transcription
The MotoTech Manufacturing Company: Design of Experiments
Additional information, including supplemental material and rights and permission policies, is available at http://ite.pubs.informs.org. Vol. 10, No. 2, January 2010, pp. 95–97 issn 1532-0545 10 1002 0097 informs ® doi 10.1287/ited.1090.0041tn-b © 2010 INFORMS I N F O R M S Transactions on Education Teaching Note The MotoTech Manufacturing Company: Design of Experiments/ANOVA Prakash Mirchandani Katz Graduate School of Business, University of Pittsburgh, 358 Mervis Hall, Pittsburgh, Pennsylvania 15260, [email protected] Key words: design of experiments; ANOVA; process improvement; interaction effect; main effects History: Received: June 2009; accepted: November 2009. Distribution. To maintain the integrity and usefulness of cases published in ITE, distribution of these teaching notes to any other party is prohibited. Please refer interested instructors to ITE for access to the teaching notes. Part A. Do you agree with Nadine’s basic approach outlined in Figure 1? Under what conditions will this approach work? Under what conditions will it not work and, therefore, have to be modified? Nadine has adopted a systematic approach for determining whether the two factors (time and temperature in the example in Figure 1) being studied affect the product quality. The approach first identifies an approximate location for the best combination of the two factors and then zooms in to locate this combination more exactly. For this approach to work, several conditions must hold. These conditions are described below. If these conditions do not hold, then we might end up searching in the “wrong” neighborhood, ending up with a process setting that is a local, but not a global, optimum. First, the initial range of factors chosen for firststage experimentation must be wide enough. If this is not so, then the neighborhood identified in the first stage might not be close to the neighborhood that does contain the best combination. Therefore, we might end up looking in the wrong neighborhood. No amount of fine-tuned search will help if this case. For example, if the best combination is 40 F and 150 minutes, the approach would not work because the initial range chosen by Nadine does not include this combination. However, the question then becomes: How does one determine the initial range within which to experiment? One possible way for determining the appropriate range to use in Phase 1 is to fall back on the extant knowledge and experience that MM’s employees have. Also, if the optimal combination turns out to be near a boundary of the test region, then the boundary may need to be expanded before one zooming in begins. Second, the relationship between the input factors and the output factor must satisfy some properties. Suppose the relationship between the response variable “quality” and the input factors “time” and “temperature” is given by quality = f (time, temperature). A necessary condition for the approach outlined by Nadine to work is that a local maximum of f is also a global maximum. (Loosely speaking, we want f to be unimodal, or tent-shaped, or if it is multimodal, each maximum solution value is identical. If f has multiple global optima, the approach will still work by identifying one of the optimal solutions.) If this condition is not satisfied, then the first stage of the investigation might identify a time-and-temperature combination close to a local optimum, and the second stage might just identify the local optimum. For example, for a response function in Figure TN1(a), the approach will work, but in the case of Figure TN1(b), the best first stage combination might be at the one of the two corner points, and the approach will identify the corner point as the global optimum, whereas the global optimum is in the middle of the factor ranges used. For the approach to work, we also have to ensure that the other factors that could possibly affect quality are also considered. For example, suppose barometric 95 Mirchandani: Teaching Note: The MotoTech Manufacturing Company: Design of Experiments/ANOVA Additional information, including supplemental material and rights and permission policies, is available at http://ite.pubs.informs.org. 96 INFORMS Transactions on Education 10(2), pp. 95–97, © 2010 INFORMS Figure TN1 Example of Situations Where the Approach May Work (Figure TN1(a)) and Where It May Not (Figure TN1(b)) (a) Unimodal relationship 55 60 65 1.0 0.5 0.0 115 110 105 100 (b) General relationship 1.0 0.5 0.0 55 120 115 60 110 105 106 65 pressure and supplier category are two additional factors. MM may know through prior investigation that these factors are not important, or these factors may not be controllable by MM, in which case MM may want to try to keep them constant as much as possible. If this is not so, MM may want to include these two factors and conduct a higher (four, in this case, because we have four factors) dimensional experimental design. Another factor to keep in mind is the measurement scale of the input factors. The approach in Figure 1 (of the case) assumes that both factors are continuous variables, so we can continue to successively fine tune the settings. When the factors are categorical, one can still do stage 1 of the analysis depicted in Figure 1, but the stage 2 analysis may not be feasible, or may not be needed, because factor categories may not be further divisible. Part B. What does John Tagole mean when he says that some other department can become the bottleneck department? The diffusion department has been a bottleneck until now, because the quality problems in the diffusion stage have been limiting the quality that MM can offer its customers. As the quality of the diffusion stage improves, and the quality of the processes that precede diffusion or those that follow is kept unchanged, one of these other departments might start limiting the quality that MM can deliver to its customers. Once this happens, the bottleneck shifts from the diffusion stage to another stage of the process. At that point, quality issues in these other departments would need to be addressed if MM wants to further improve the quality of its products. Part C. For a tangible product or service product that you are familiar with, briefly describe three CTQ dimensions. Some CTQ dimensions for a call center are Time on hold; Accuracy of information provided; Repeat calls for the same problem. Some CTQ dimensions for an LCD television are: Time to failure; Quality of the picture; Number of dead pixels. Part D. Do you agree with VR4U’s recommendation that supplier and temperature do not affect thickness? If yes, why? If no, why not? What would you recommend? VR4U has done a univariate ANOVA analysis, analyzing the supplier and temperature factors separately. From this analysis, we conclude that neither factor is significant. (Please see MotoTech (DOE) Solution.xls spreadsheet.) However, what we need here is two-way ANOVA. To do this, we have to reorganize the data first. The null and the alternate hypotheses for checking the interaction effect are as follows: H0 : Supplier and temperature do not interact to affect the thickness. HA : Supplier and temperature do interact to affect the thickness. Looking at the ANOVA results, we see that the p-value of the interaction effect is less than 0.05 (Table TN1). Therefore, we can reject the null hypothesis at a significant value of = 001 and thus conclude that supplier and temperature do interact to affect the diffusion thickness. Because there is an interaction effect, there is no need to do the main effect tests. Looking at the plot of the means (Figure TN2), we can understand why we see the main effect is not significant and the interaction effect is. If we average across each supplier (ignoring the temperature), then the mean for each supplier turns out to be about 3,010 Å.1 Similarly, if we average across each temperature level (ignoring the supplier), then the mean for each temperature level turns again out to be 3,010 Å. The graph shows that the thickness is affected by the 1 An angstrom (Å) is a unit for measuring length and equals 10−10 meters. Mirchandani: Teaching Note: The MotoTech Manufacturing Company: Design of Experiments/ANOVA 97 Additional information, including supplemental material and rights and permission policies, is available at http://ite.pubs.informs.org. INFORMS Transactions on Education 10(2), pp. 95–97, © 2010 INFORMS Table TN1 ANOVA: Two-Factor with Replication Summary High Low Medium Total Pinnacle Count Sum Average Variance 15 4523534 3015689 4.787305 15 4507834 3005223 6.163007 15 4514254 3009503 1.624787 45 1354562 3010138 22.88096 Allied Count Sum Average Variance 15 4515179 301012 2.676932 15 4514744 3009829 3.333901 15 4515075 301005 4.292004 45 135450 3010 3.293833 Premier Count Sum Average Variance 15 4507187 3004792 3.120144 15 4522643 3015095 6.922027 15 4516602 3011068 3.4528 45 1354643 3010318 22.67741 45 1354522 3010049 21.86212 45 1354593 3010207 3.411622 Total Count Sum Average Variance 45 135459 30102 23.61436 ANOVA Source of variation Sample Columns Interaction Within Total SS df MS F P -value F crit 2.297154 0.717594 1,639.559 5092207 2 2 4 126 1148577 0358797 4098896 4041434 02842 008878 1014218 0753097 0915104 1.99E−38 3.0681 3.0681 2.443591 2,151.794 134 combination of the supplier and temperature levels. Because, there is an interaction effect, there is no need to do the tests for the main effect. What should John Tagole do? If the temperature can be held steady at high, he should select Premier, or if the temperature can be held steady at low, he should select Pinnacle. If the temperature can be held steady only at medium, the choice of supplier does not significantly affect the thickness. If the temperature cannot be held steady, it makes sense to buy from Figure TN2 Plot of Means 3,020 3,015 3,010 3,005 Low 3,000 Medium High 2,995 Pinnacle Allied Premier Allied, as the output measure when Allied is used does not vary with temperature. Depending on the sophistication level of the students, instructors may also want to discuss the concept of robustness. A robustness perspective will argue for selecting Allied because the output when Allied is the supplier is insensitive to temperature. Even if temperature can be held relatively steady at the high or low levels, it might exhibit some variability and thus result in poor quality for the Premier and Pinnacle cases respectively. Note from Table TN1 that at low temperature, Pinnacle’s mean is closer to the target value than Allied’s is and at high temperature Premier’s mean is closer to the target value that Allied’s. For all three temperature levels, though, the process mean for Allied is higher than the target value. If Allied is chosen, MotoTech and Allied should work together to bring the process mean closer to the target value. Generally, it is easier to change the process mean (which might simply require the operator to be better trained) than to reduce the process variance (which might require investments in new technology). Supplementary Material Files that accompany this paper can be found and downloaded from http://ite.pubs.informs.org.