Design of experiments helps reduce injection molding scrap rate by 90%.

Robinson Rubber Products, Minneapolis, MN, previously experienced an intermittent scrap rate of up to 10% while producing a challenging injection molded part involving an inner and outer steel sleeve with molded rubber bonded in between. The most common problem was that the part did not pass the company's tough metal-to-rubber bonding test, which requires that the rubber actually tears apart and does not pull away from the metal. Over the years, Robinson made many one-factor-at-a-time (OFAT) changes to the process, but the problem continued to reoccur.

Robinson's quality assurance manager worked with Stat-Ease, Minneapolis, MN, a vendor of Design of Experiments (DOE) software, to perform an experiment that evaluated the effects of five key manufacturing variables. The experiment determined the value for each variable that optimized the rubber-to-metal bond. Now that these values are consistently used, the scrap rate has stayed well below 1%, with no recurrence of bond problems.

Robinson Rubber Products designs and manufactures custom-molded rubber components, extruded profiles and precision rollers for original equipment manufacturers and industrial companies worldwide. The company specializes in complex components, including multiple insert parts, homogenous rubber parts, and rubber bonded to metal and other substrates. Robinson's major markets include automotive, off-highway and heavy-duty vehicles, machinery and material handling equipment, home and commercial appliances, and many other industries.

Tough rubber-to-metal bonding problem

The part mentioned above uses two coats of a bonding agent applied to each of the metal inserts. The inserts are hand-loaded into a 16-cavity mold, the rubber is injected, and the parts dwell in the mold for a specified time, temperature and pressure. To verify the bonding, operators periodically perform a destructive test on the product by pushing the inner steel insert out of an assembly and visually checking the rubber-to-metal bond.

In the past, the company would periodically experience a large number of rejects because of bond failure. Engineers and operators having experience with the process would each express their opinion on the problem. It was easy to disagree, because the manufacturing process for this part has many variables. The injection speed, injection pressure, vulcanizing temperature and dwell time of the injection molding machine can easily be adjusted. The thickness of the bonding agent and amount of time it is allowed to dry are other important variables. Process changes would be made on the fly until the problem went away. Then, after a period of time, the problem would reappear and the trial and error process would start all over. When this happened three times in only two weeks, the firm knew that it needed to find a new method to find the root cause of the problem.

A major problem with the OFAT method is that it cannot detect the interactions of multiple factors. By varying an individual variable, one can find the optimal value of each one, with all the others held constant. However, when one combines the supposedly-optimized values of each variable, the results are usually far less than optimal, often because of the ways that they interact with each other.

Design of experiments evaluates interaction of variables

Past experience with similar problems suggested that the only way to really understand what was happening was to perform a designed experiment to find the critical factors that are required to achieve a robust rubber-to-metal bond. DOE provides a better approach that varies the values of all variables in parallel, so it uncovers not just the main effects of each variable, but also the interactions between the variables. This approach makes it possible to identify the optimal values for all variables in combination, and also requires far fewer experimental runs than the OFAT approach.

Experimental designs are available with as little as k+1 runs, where k equals the number of variables to be tested; for example, seven factors in eight runs. But more accurate, results can be achieved by higher resolution designs which are capable of evaluating the main effects of each variable, as well as the two-factor interactions. For example, 16 runs suffice for a high-resolution design on five factors, only half of the 32 combinations of all these factors at two levels each.

Stat-Ease was selected to train the firm in design of experiments and assist with the experimental set-up and analysis. Stat-Ease pioneered the development of personal computer-based design of experiments software that simplifies the process of solving real-world industrial problems. The company's software package, Design-Expert, walks users through the process of designing and running the experiment and evaluating the results. A Stat-Ease engineer presented a four-hour training session on DOE to Robinson Rubber staff, and then everyone worked together to select the factors that they thought were the most likely to cause the bond problems. Robinson's quality assurance manager was chosen to coordinate and monitor each step of the experiment.

Experimental design

The five factors that were selected were:

* Vulcanizing temperature (low versus high);

* bond material (two types);

* bond application thickness (one coat versus two);

* injection pressure (low versus high); and

* bond settling time (low versus high).

Robinson Rubber engineers entered this information into Design-Expert, and the software generated an ideal two-level design with 16 runs.

The experiment was run on the first shift over two days. The design blocked out the day-by-day impact. It included an additional six cures at the beginning, middle and end of each day, with the same conditions to serve as controls to detect the effect of any variables besides those that were being measured in the experiment. The order of the cures was randomized by Design-Expert as insurance against lurking variables, such as machine wear. Slight modifications by the quality assurance manager were introduced to simplify the temperature changes. Every cure was monitored in person by quality assurance technicians to ensure that all the factors were correct for each cure.

Three different tests were used to evaluate the results. The force required to push the inner steel core out of the part was measured in tons. The amount of rubber bond coverage remaining on the steel sleeve was measured on a scale of one to 10, with one being almost no bond and 10 being 90% to 100% bond remaining. A rating of eight or better was considered to be an acceptable bond in accordance with the drawing specification. The completeness of the cure was rated on a one to five scale, with one being a very low state of cure and five being completely cured. A rating of three or better was considered acceptable.

Optimizing the values of each variable

Quality assurance technicians did the push-out testing and rated the parts together. All the test results were delivered back to the Stat-Ease consultant, who entered them into the software and generated reports, including an analysis of variance (ANOVA) table that identifies the factors that significantly affect the test results. The results showed that the bond settling and the injection pressure had no effect on the results.

A graphical optimization overlay indicated the optimal value of each variable, with its interactions with the other variables taken into account. Bonding agent A was clearly superior to bonding agent B. High and medium temperatures worked better than low temperatures. Finally, a thick coating of the bonding agent was better than a thin coat.

Robinson Rubber began consistently producing the part under these conditions. The scrap rate immediately dropped to well below 1% and has stayed at that low level ever since. Design of experiments is clearly a useful tool that can help determine the root cause of and solution for difficult problems that might fester for years otherwise.

by John Engler, Robinson Rubber, Jerry Fireman, Structured Information (www.strucinfo.com)