Printer Friendly

Using design of experiment to uncover process mysteries.

In one month, Waupaca Foundry officials discovered valuable data after combining three manufacturing operations into 32 experiments--in what otherwise would've required 1024 tests.

To improve the machinability of their castings at one of their customers, Waupaca Foundry officials recently suggested that its quality team design a new method of experiment. Over the previous year, the foundry had been running one-variable-at-a-time experiments, where a single chemical property or cleaning method would be changed and castings were sent to the customer to be machined. The results of these testing methods proved to be insignificant and no root cause for problems was determined.

On one particular experiment on the tool life of ceramic inserts, such tests were time consuming and didn't offer insight on how variables other than the casting itself influenced the tool life. Because of this, a team comprised of representatives from Waupaca Foundry, the machine source and the ceramic inserts suppliers assembled to brainstorm on possible variables. They targeted variables that may be influencing the number of castings achieved prior to the ceramic inserts' failure to produce castings with a root mean square value (RMS) under 80.

This article deals not so much with the results of the designed experiment, but rather on how a large experimental design (such as this L-32 design) involving three manufacturing operations was coordinated and performed in one month.

Putting Heads Together

The first step was assembling the team members to select the variables from which to run the designed experiment. In this brainstorming session, each member discussed his/her feelings regarding the possible root causes that were influencing the number of castings achieved per ceramic insert. All suggestions were recorded on a Fishbone diagram and reviewed by the entire team.
Table 1. The 10 Variables Chosen for Waupaca's Experimental
Variable Level 1 Level 2
A-Finish stock 0.090 in. 0.130 in.
B-Feed 0.006 in. 0.013 in.
C-Speed 412 rpm 578 rpm
D-Alloy Tin 1 oz/100 lb Base iron
E-Inoculation 2 oz/100 lb 4 oz/100 lb
F-Cooling time Normal Longer
G-Nose radius ofinsert 0.093 in. 0.125 in.
H-Insert hole Without With
I-Insert hone 0.0005 in. 0.0015 in.
J-Coolant Without With
Table 2. Rows Assigned to Each Variable.
Row No. Main Effects Two-Way Interaction
1 A
2 B
4 C
7 DE
8 D
10 BD+FJ
11 CE
12 CD+IJ
13 BE
14 AE
15 E
16 F
19 G
21 H EJ
22 I
24 BJ+DF
25 EI
26 J
27 AJ+DG
28 EG
29 DH
30 CJ+DI
31 EF

After completing the Fishbone diagram, there were more than 30 possible variables from which to run the experiments. Realizing the list of possible variables needed to be trimmed to achieve a realistic and controllable experimental design, team members voted for the three variables they felt were most significant and created a pareto chart from the data. By doing this simple and democratic action, the team trimmed the list to 10 variables for the experimental design.

Table 1 lists the 10 selected variables with two levels each chosen by the team for Waupaca's experimental design.

After establishing which variables to examine, the team studied which orthogonal array to use. Each of these variables needed to be a main effect variable without the "noise" created from possible two-way interaction, but the number of total experiments needed to be an achievable number. The team chose an L-32 design, meaning 32 different experiments would need to be performed.

Slotting Variables

After choosing the correct array, a great deal of care and planning must be taken when placing or assigning each of the variables to a row in the orthogonal array. Certain rows are designated for TABULAR DATA OMITTED main effect variables, while others are designated for possible two-way interactions between variables that generally turn out to be a significant part of the problem.

Table 2 represents the rows assigned to the variables. Rows 1, 2, 4, 8, 15, 16, 19 and 22 are where the main effect variables were placed, meaning if the variable assigned to that row is found to have a significant effect on the number of castings per insert, it will be entirely due to that variable--without an interaction between variables.

Two of the main effect variables in a row are designated for a two-way interaction--rows 21 and 26. The only danger here is that if either or both of these rows are found to be significant, then it must be determined whether it was due to the main effect variable assigned to that row or to the two-way interaction that was designated for that row on the orthogonal array. In a situation like this (where you need to place a main effect variable into an interaction row), it is extremely important to choose a row or rows that won't have an interaction between the variables that were designated for that row.

It is recommended to use no more than one extra main effect variable on an L-8 or smaller design, and two to three extra main effect variables on L-16 and L-32 designs.

Preparing for the Test

After completing the preliminary work, it's time to start preparing the castings and ceramic inserts for each of the 32 experiments. This isn't as difficult as it may seem, as you can see from Table 3. To fulfill Waupaca's requirements for the design of experiment (DOE), the team only needed to be concerned with variables A, D, E, and F--those applicable to the foundry.

Table 3 is the final orthogonal array used to perform the experiment. Due to the size of the experiment (32 total tests), officials decided to run eight tests per day for four days. By doing this, they were able to arrange the orthogonal array into four blocks, representing possible influences due to the variations in the machining equipment, setup and environmental changes that could affect the performance of the machining experiment.

This was done by changing the order in which each test was run, as long as when the final results of each test were entered into the software program, the order of experiments is the same as called for by the orthogonal array.

Incidentally, in Table 3, only the columns that represent the main effect variables are used to describe how to perform each test. You do not need to address the two-way interactions until you analyze the final results of the software.

In this case, 150 castings needed to be produced for each of the 32 experiments, although typically it isn't necessary to make that many parts per experiment. Because the number of castings per insert needed to be evaluated, however, team members had to guarantee there would be an adequate number of castings per experiment if they were to hit the correct combination of variables, giving an insert life number of over 100 castings. That was the goal at the onset of the proposed DOE.

If you look at experiment Number 1 on Table 3 under variables A, D, E and F, you'll see how the team determined the methods used to produce castings for experiment No. 1. The foundry produced 150 castings as close as possible to this list of requirements.

If you look at experiment No. 5 on Table 3, it has exactly the same requirements for the foundry variables as No. 1. The only difference between experiment No. 1 and No. 5 is that the levels of the ceramic inserts and machining operations change.

This pattern holds true for all 32 experiments, allowing officials to run only 16 experiments of 300 castings each. This saves valuable setup time and produces the castings faster and more uniformly.

Finishing Touches

For the first run, 300 castings were produced. The first 150 pieces were tagged "Experiment No. 1" and the remaining 150 pieces "Experiment No. 5." There were no rules stating which order to run the experiments, as long as they were identified correctly.

For instance, the first variable, "A," is finished stock, which requires the foundry to produce castings at both 0.090-in. and 0.130-in. finish material. This required adding 0.040-in. shim stock to both brake plates in order to get the 0.130-in. finish required for Level 2. This was very difficult to control, since it was time consuming and the team didn't want to add or remove the shim stock any more than necessary.

The team concluded they should produce all eight of the 16 experiments that required the extra stock first, then remove the shim stock and produce the final eight experiments.

The eight experiments to be run first with 300 castings each (150 for each experiment) were as follows: No. 1 and No. 5; No. 3 and No. 7; No. 9 and No. 16; No. 11 and No. 15; No. 17 and No. 23; No. 18 and No. 24; No. 25 and No. 32; and No. 27 and No. 29.

The second hardest variable to control was "F"--the cooling time. For this set of eight experiments at 0.130-in. finish stock, the team produced the four tests that required the normal cooling first and then the second four tests that required longer cooling last. Table 4 lists the order the first eight experiments ran. Regardless of which order you run the experiments, each must be tagged or identified according to the number assigned to that experiment on the orthogonal array.

The same theory applies to the order of the second set of eight experiments. By taking this extra time to set up the order for each test lot, the team produced TABULAR DATA OMITTED all of the castings required during normal production in two days, with very little productivity lost. To pull this off, much time was spent coordinating every department's responsibilities before producing each test lot. With adequate planning and cooperation, such an endeavor can be completed without the personnel involved pulling their hair out in frustration.

After the castings were produced and tagged correctly, they were packaged and sent to the foundry's customer. In addition, the ceramic inserts were produced and sent to a mutual customer. The experiments were performed at the machining source in a four-day period, with assistance from representatives from each of the involved manufacturing facilities. The number of castings per insert was determined by the RMS number on each brake plate achieved until failure of the insert or violation of the specification.

DOE Gives Insight

The results from each of the 32 experiments were then analyzed using the ANOVA table on a software program. The variables that showed to have a significant influence on the ceramic insert's life are listed in Table 5.
Table 5. Results: Variables Influencing Ceramic Insert's Life.
Variable Best Level
A-Finish Stock 0.130 in.
B-Feed 0.006 in.
D-Alloy 1 oz Tin/100 lb
BD-Feed X Alloy 0.006 in. at 1 oz Tin/100 lb
AG-Finish Stock X Nose Radius 0.130 in. at 0.125 in.
G-Nose Radius 0.125 in.
BJ-Feed X Coolant 0.006 in. at no coolant
J-Coolant None
AJ-Finish Stock X Coolant 0.130 in. at no coolant

The results proved to be exciting, but confirmation runs must always be performed at the recommended levels to determine the experiment's true value or significance. In this case, the confirmation runs verified the results of the DOE. By doing this, Waupaca was extremely confident that once implemented, the recommended levels of each significant variable would increase the number of castings per insert in a production mode.

The other item not to forget, which may result in great cost savings, is analyzing the variables that were found to be insignificant and producing the product associated with that variable at the more economical of the two levels chosen for the DOE. For instance, the team discovered cooling time was insignificant, which means Waupaca could produce the castings at a faster cycle time for improved productivity.

Waupaca previously conducted years of one-at-a-time experiments with no significant results. If the foundry continued performing one-variable-at-a-time experiments, it would've needed to run 1024 more experiments to achieve all possible combinations. The entire process, from initial planning through the final implementation of the results, was performed in fewer than 30 days.

Using the fractional factural method, the foundry achieved all combinations in 32 experiments. A well-performed fractional factural designed experiment method for problem solving is a powerful way to identify and improve your foundry's operations.
COPYRIGHT 1993 American Foundry Society, Inc.
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1993, Gale Group. All rights reserved. Gale Group is a Thomson Corporation Company.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:Waupaca Foundry Inc.
Author:Ebert, Steve
Publication:Modern Casting
Date:Jun 1, 1993
Previous Article:Controlling melt components can lower good casting costs.
Next Article:Economics of gunned plastic refractories gaining interest.

Related Articles
Technology continues to evolve.
Purchaser's notebook: how to buy the right core machine.
Wisconsin group examines new techniques, basic methods.
Recycled steel can scrap is a proven iron foundry charge.
Waupaca breaks out of the mold with new gating approach.
Charge material issues examined by ferrous foundrymen.
The year in review.
Using Eddy current testing to identify casting defects.
What variables affect core erosion resistance?
Iron & Steel MACT: AFS hosts EPA for plant tours to educate officials. (Washington Alert).

Terms of use | Copyright © 2016 Farlex, Inc. | Feedback | For webmasters