Sir Isaac Newton was right. Every action begets a reaction. In the highly processed world of steel manufacturing, knowing how every event can effect subsequent events is of paramount importance since processing line shutdowns are very costly.
Several years ago Dofasco Inc., the giant Hamilton, Ont., steel company decided to take the tremendous amount of data gathered in the steel-making process and put it to use (real-time) on the plant floor.
Dofasco’s Michael Dudzic said, he and his fellow workers, the advanced control, data modelling group, are always looking at control and modelling solutions, and analysis and statistical technologies to help improve the ability to predict manufacturing breakdowns. Like other companies, Dofasco is gathering a great deal of data through constant monitoring of the steel-making process, he said.
We wanted to maximize the value of all this data, so it came down to finding a way to get the best results from the data, Dudzic, manager of process automation technology, said.
That is essentially the crux of the problem. How can a company take millions of bits of data and get useful, predictive information that can be taken live, onto the plant floor? The solution requires the need for plant operators to be able to act in a proactive manner rather than reactive. Solving the problem after the fact is of little use.
Though most large manufacturing fields, such as pulp and paper, mining, oil and gas, and petro-chemical, use statistical models to help solve manufacturing problems, Dofasco’s needs were specific, in part, since it was dealing with the very high temperatures associated with molten steel.
working together
For Dofasco, the solution came from its long standing relationship with Hamilton’s McMaster University. About 15 years ago, the university’s chemical engineering department set up an advanced control consortium, with about 20 large manufacturing companies on board, with the intention of looking at using solutions such as advanced multivariate statistics to solve manufacturing problems.
A Cornell University tutorial best describes multivariate statistics as, “techniques (to) look at the pattern of relationships between several variables simultaneously.”
John MacGregor, professor of chemical engineering at McMaster University in Hamilton, said it is not uncommon for something like a chemical refinery to measure thousand of variables every second. With more computer use, there was more data.
“So they really just created these data graveyards, where they dumped the data and they never appeared again,” he said.
Robin Wurl, a professor of industrial and manufacturing engineering at Oregon State University in Corvallis Or., agreed with MacGregor.
“There is a sensor on every single piece of the process and it is spewing data,” she said.
Traditional statistical methods, based on smaller data sets with well defined assumptions, tend to fall apart with large data sets, MacGregor said. In a situation like Dofasco’s, though there are thousands of variables, there are only about three or four things happening in the process.
“The data is very highly correlated, there are usually five or six key events that could be extracted and many of the other data are correlations or other instances of those key events,” Dudzic said.
“The power of this technology is that it assumes that there is going to be a lot of correlation in the data and it just looks for all the unique aspects,” he added.
“You get a tremendous compression with multivariate statistics, you take this thousand dimensional space into this low dimensional space,” MacGregor explained.
This allows Dofasco’s Caster operators to monitor the entire process, while only having to look at a limited number of factors. “What we see from the operators perspective is that it just gives them a bit more of a comfort (zone),” Dudzic said.
For example, in the steel making process precise temperatures are extremely important. If the steel is coming through and a side or corner of the molten slab is not properly formed then the relationship between all of the temperatures would be disrupted, MacGregor explained.
Though many of the different temperature gauges might read within the acceptable norm, the relationship between all of the temperatures would be thrown out of kilter if one reading was out of the norm. So instead of having to monitor all the temperatures, the operator can monitor one relational temperature. If the alarm sounds that the temperature is trending away from the norm then the operator can dig down to find where the problem is occurring and rectify it before it becomes a real problem, in this case a breakout; when molten steel breaks out of the mould and causes damage to the machinery.
“This tool along with all of the improvements that have gone on, we have had very significant improvement overall in our performance in breakout reduction,” Dudzic said.
Though the total cost and time of implementing the multivariate statistical solution is confidential, Dofasco plans on bringing the same technology online to their number two Caster, sometime in the future.
