When it comes to analyzing, interpreting and using data, your point of view can make a big difference. Those of us who spend most of our time in controlled laboratory environments can easily forget how chaotic the "real world" can be. Variation in process parameters, raw materials and human interactions can skew the picture of what we're studying. To be effective, we need to understand the data in context and be on the lookout for unusual variation.

Lynda Hushard, Quality Assurance Supervisor for Huntsman Corporation, a chemical manufacturer in Guelph, Ontario, couldn't agree more. "Some problems have a clear-cut cause and effect," she said. "But there are cases when a non-spec test result can be caused by 10 or 15 different variables, or—even more difficult—variation that you're not testing for. When that happens, we use statistical analysis to graphically 'see' the process and learn as much as we can about the variation that's occurring."

Huntsman's Ontario plant makes more than 350 different products, all by batch process methods. With 25 kettles going simultaneously, the QA lab manages a large amount of data. Typically, operators pull samples from the kettles at various stages in the process and deliver them to the lab. Lab technicians analyze the samples and record the data in a LIMS database. This process data can be exported to statistical software for further analysis.

Hushard says that data management and analysis is a critical part of what the lab does, because it creates a history of the process and is invaluable for solving problems. The majority of the lab's data management involves collecting and analyzing test data for charting trends on factors such as product, kettle and time. "If we get a sample that shows a trend swing, we can go back into the data, check dates and compare it with supplier changes, deliveries of new loads of raw material," Hushard said. Other data management involves variability testing on the lab's methods and instruments, as well as reproducibility studies.

In a recent situation, one of the plant's regular products, a surfactant, began showing substantial variation in color, producing off-spec product. Further testing data showed that in subsequent batches the color came back down to normal, then intensified again.

Hushard and the plant's staff went through a process of elimination in analyzing production variables to discover the source of the color variation. "It was an interesting problem," she said. "At least five different variables, or a combination of them, could have caused the color variation."

Nothing conclusive turned up until they used control charts with dates plotted at the bottom. Control charts show the extent to which measurable characteristics relevant to the process are in-control (statistically predictable) or out-of-control.

Next they overlaid the chart with supplier change dates. "And what do you know," said Hushard, "the raw material delivery from one supplier lined right up with the color changes. It showed us in no uncertain terms that we'd gotten several lots of raw materials with a bad color."

Hushard said the control charts allowed them to see unusual variation in the process that they don't usually look for in the course of normal laboratory testing. "Since color wasn't a typical specification for that particular ingredient, our normal testing protocols didn't catch it," said Hushard. "But thanks to the ease with which we were able to produce control charts and overlay them with the delivery schedules, we were able to pinpoint the problem and get back on track. It was very powerful information."

Analyzing Flow Rate Variation

Tosco is one of the largest independent oil refining companies in the U.S., converting crude oil and other feedstocks into finished petroleum products such as gasoline and diesel fuel. The Royal Grande plant handles an intermediary step in the gasoline-making process, refining and upgrading 44,000 barrels of crude oil per day.

Davis's job is to analyze flow rates of crude oil going in and products coming out of the plant, with the goal of maximizing production of higher quality petroleum products. Crude oil has millions of different chemical compounds, some of which are higher-value gasoline and most of which are lower-value large molecules such as asphalt. The initial refining process separates out gasoline and diesel fuel, then the large molecules in the remainder are broken up into smaller molecule compounds (e.g., diesel, gasoline, sulphur and coke) by heating them to 900 degrees in a coke drum.

During the refining process, hundreds of different parameters, such as temperature and pressure, can affect which products are produced. Davis' group tracks how well the process is working by analyzing flow rates for different products that the process is producing.

Flow rate data is collected from Rosemont orifice meters in the pipes and is sent to a Rosemont RMV9000 distributed control system (DCS) for storage. Each day, a Digital Equipment Corporation VAX computer extracts 24 hours worth of data from the DCS and downloads it into Microsoft Excel. From there it is automatically transferred into Quality Analyst from Northwest Analytical for control charting.

Figure 1

A typical control chart that is used to keep accounting measurements in agreement with field metering. The field meters drift with crude viscosity and have to be adjusted regularly. Note that near the right side of the upper chart, several points in a row are slightly above the center line. This indicates that the field meter is reading slightly too high and needs adjustment.

ENLARGE IMAGE

Davis primarily uses running average and range control charts for analysis. The software is configured to automatically adjust for amounts of crude coming into the system. When the control charts indicate something is amiss, a detailed evaluation of the other variables ensues, and the team works to restore optimal performance.

Figure 2

These charts compare gasoline yields on two identical plants. Note that yield on the left side is drifting down over time while yield on the right side is drifting up. This situation would call for a process evaluation to find out why one plant is not doing as well as the other.

ENLARGE IMAGE

Davis said that the key to understanding the data is to work closely with the operators. "If the data goes out of control limits, I go right to the control room and talk with them," he said. "Three out five times, the operators can point me to the cause of the problem. Combining their hands-on process knowledge with the data analysis gives us a very clear picture of what's going on."

For example, Davis occasionally contends with variation in data from the orifice meters that measure flow rates for the main crude feeds. "The meters are sensitive to changes in viscosity, which can vary a lot with crude oil," he said. "There's no way for an operator to know how much the viscosity is varying. We deal with it by running control charts, looking at the data trend, and when the data lines up in a row above or below the center line on the control chart, we know there's something consistently wrong with the flow rate and can adjust the throughput."

Ease of Analysis is Key

Davis and Hushard both agreed that having the right tools to collect and manage data is key to understanding and dealing with unusual variation. Hushard said that performing statistical analysis on her lab's own testing methods showed that human factors in manually weighing, diluting and injecting samples created substantial variation in the results.

"With six people in the lab, we had a lot of natural variation in doing manual injections of samples into the testing instruments," she said. "When you're dealing with microliter volumes, simple things such as whether you close one eye or leave both open can affect whether you're injecting at a perfect 90 degree angle. So you end up working a larger and larger hole in the septum."

After analyzing the lab's process with control charts, she was able to justify the purchase of a PerkinElmer autosampler. The variance dropped from five percent to less than two. "We brought the variance down more than 50 percent," Hushard said. "With far more consistent results, we have a lot more confidence in our sampling methods."

What makes both Davis and Hushard successful in their work is knowing how to see beyond the data points in front of them. A good understanding of the process, willingness to work with the operators and above all, a solid understanding of variation is what makes the difference for them. The lesson is clear: variation may be the spice of life, but like most spices, a little goes a long way. Find the unusual variation, and you'll be much closer to reaching quality and production goals.

Jeffery L. Cawley is vice president of Northwest Analytical, Inc. He can be reached at jcawley@nwasoft.com. 

back to top