Archer Daniels Midland (ADM) operates more than 760 agricultural products plants around the world, using corn, soybeans, wheat, oilseeds and other produce to manufacture materials used to make products ranging from food and animal feed to biodiesel and industrial feedstocks.
In 2004, the company recognized the potential for plant data to help improve performance in regulatory compliance, quality, efficiency, process design, cost accounting and logistics, and launched an initiative to improve its ability to convert data into useful information. "Over the past 11 years, we've made some progress," said Scott Harmeier, process optimization manager at ADM, to attendees at the NovaTech Futureproof Automation 2015 User Conference on Aug. 31 in New Orleans. "It's been a long and constantly changing process."
Harmeier shared key insights ADM has gained through the experience. "The first step was to select and install a historian in our corn plants," he said. "We hired a database administrator for each plant, who implemented the technology, but did not understand how to make the data valuable to the process. We then gave the data to the process engineers, who were able to make significant, but intermittent and isolated improvements."
Encouraged but not satisfied, in 2011, ADM introduced Operational Excellence, which is intended to drive continuous improvement. "We ran into some issues," Harmeier said. Using a DCS tag-based method to find data was OK for a single plant, but became a problem when trying to transfer an improvement to multiple plants with different control systems and tag bases.
"It's also possible to calculate KPIs in the DCS, but it's hard to do on the scale and in the formats needed for reporting. For example, we had trouble shifting time scales from once-per-second DCS data to the averages over minutes, hours and days needed for reporting," Harmeier said. Spreadsheets are "OK, but hard to scale up. It takes too long to do the calculations."
Databases are good for collecting and historicizing data, but converting it into information requires ways to find, analyze, deliver and visualize it. To help find the data, ADM adopted the concept of an asset framework. "Instead of a tag-based search, we built a hierarchy of asset structures, so similar equipment or assets in different facilities have the same data structure," Harmeier said. "We created templates for each asset type, and use them to drive server-based analytics in real time."
The analytics include methods such as regression, decision trees, neural networks and clustering. "With the two groups involved—the process engineer who understands the process, and the data scientist with the skills to run the algorithms—they can work through root causes and ultimately solve a problem," Harmeier said. Then the asset structure allows the data scientist to take that solution to similar situations across the enterprise.
A similar collaborative approach is needed to bridge the data silos associated with different departments and facilities. "Big data is big because of its volume, velocity, variability, veracity and variety," Harmeier said. "Within a silo, the experts can deal with the first four Vs. Across the company, the key is dealing with the variety of data types."
Integrating information across silos and improving communication and collaboration between siloed groups and functions requires software that allows them to use each others' data. All the plant systems, all the users, and the data flows must connect. Multiple plants and functions, and integrate the corporate level business systems. "Avoid point solutions, invest in flexible systems, and find an enterprise architecture that can be integrated with existing systems," Harmeier said.
Now, 11 years into the experience, ADM is "still implementing and evaluating," Harmeier said. "But we're well along."