The steel group voestalpine shows how to save. Using the example of the scale error prognosis model in steel strip production, it becomes clear that standard statistical procedures are often not sufficient to recognize and explain complex contexts. The automotive industry places the highest demands on the surface quality of a steel strip. So-called (rolling) scale defects disturb considerably further processing processes and can accordingly lead to the scraping of the entire steel strip. As a result of (rolling) scale faults, the steel group voestalpine incurred costs of around one million euros per year.
Background to the production
The aim of quality assurance was therefore to develop a statistical model of the production conditions for forecasting the scale defects. From this it should be possible to deduce with which production conditions the formation of scale defects can be avoided. Higher mathematical methods and multivariate considerations as well as the consideration of different production processes led to a success and a reduction of the error costs by 80 percent.
Tasks and analyzes
At the Linz plant of the voestalpine-Division STAHL, roughly 5.5 million tonnes of crude steel are shed each year to form slabs. In order to be able to produce sheet metal from this 30 ton sheet metal for further processing in automotive production, the steel has to be rolled to the desired thickness into steel strips. In (hot) rolling, under certain conditions, the scaling layer is rolled into the base material on the surface of the steel strip. This defect is called the Walzzunderfehler. Lighter defects on the steel strip can be removed at the pickling lines.
Significance of results
In the furnaces, the slab is heated by direct firing with gas to around 1200 degrees Celsius. In the roughing stage, the slab is reversed to form a preliminary strip. Thereafter, the pre-strip in the finishing line is continuously rolled to the desired final thickness dimensions. Finally, the finished steel strip is cooled and reeled to temperatures between 550 and 740 degrees.
Standard procedures vs. Holistic process modeling
Following the finishing line, the steel strip is controlled by means of a surface inspection device. With this device, it is possible to detect scale defects and to assess the quality of the steel strip. Hundreds of process parameters and millions of measured values are recorded per production day and stored in databases. The size of the databases (central data warehouse with the process databases) is several terabytes.
The first question was which process parameters depend on the occurrence of walrus defect. With the help of data mining, voestalpine was able to clarify this by using higher statistical-mathematical methods. Statistical standard procedures fail miserably because of the complex relationships between production conditions and the scale errors.
One reason for the poor results, for example, of Pearson's correlation calculation are certain prerequisites, which are only rarely fulfilled - including normal distribution and linearity. Spearman's rank correlation often results in inadequate results because strictly monotonic relationships are assumed.
Two evaluation tools are used in voestalpine: STATISTICA Data Miner / Process Optimizer and voestat. In addition to the standard methods mentioned above, STATISTICA contains a variety of analysis methods, such as the nonparametric statistics and decision trees. Voestat is an efficient, standardized analysis software. The program contains complex statistical-mathematical algorithms developed by voestalpine.
The calculation using the (linear) regression or correlation computation would give a "correlation" of just five percent and thus the realization that the temperature had practically no influence on the formation of scale errors.
Voestat, on the other hand, calculated a coefficient of 0.51 - 51 percent of the scale errors depend on the temperature. The band temperature therefore has a strong influence on the formation of scale defects.
In order to be able to really predict and avoid the whirlwind errors, additional multivariate observations are indispensable. The above-mentioned questionnaire is univariate: Which process parameters depend on the flawless errors? The corresponding multivariates are complementary: Which combinations of process parameters do the faulty flaws depend on? An application example shows the problem: On the one hand, it has been found that the duration of use of the rollers alone does not have a strong, unambiguous influence on the waltz-bite error.
The results of the maximum likelihood function values in STATISTICA indicate that the variables considered combined have a significant influence. Voestat has calculated that 40% of the formation of the scale defects depends on the interaction between the duration of the application and the type of walnut. A programmatic support is crucial, because with 200 process parameters, there are already 19,900 possible combinations of two variables. For three variables, there are more than one million possible combinations.
Ultimately, STATISTICA and voestat 34 have discovered significant connections, the results of the two software packages are not necessarily identical. On this basis the prediction model was developed in STATISTICA using the "Generalized linear models" module.
The company voestalpine
The power of the model is extremely high: 95 per cent of the whale claw errors were correctly predicted. The findings of the model were implemented directly in the production process. The model is also used to monitor the production and, if necessary, alarms.
The diagram below shows the importance of the use of higher mathematical methods and the combined consideration of different process parameters and whole processes. This is because statistical standard procedures alone (model variant I) would not have allowed a substantial reduction in the error costs.
Even higher univariate statistical methods (model variant II) predict only inadequately faulty segments. However, with higher univariate and multivariate statistical procedures (model variant III), the sum could be reduced by 80 per cent. This equates to annual savings of 800,000 euros.
Voestalpine has taken its own path in Six Sigma. This is not to be a fundamental criticism of the model of quality management with its five phases. It only shows that the conventional six-sigma methodology - at least for problems at voestalpine - is based on inadequate standard statistical procedures. Furthermore, Six Sigma considers individual processes isolated and has a limited, univariate character from the outset.
Voestalpine is a successful international group with a multitude of specialized and flexible companies that manufacture, process and further develop high-quality steel products. The company group, headquartered in Linz, has around 360 production and sales companies in more than 60 countries. The former Austrian state group consists of a total of five divisions and employs nearly 40,000 people worldwide. Sales in the 2009/10 financial year totaled around EUR 8.6 billion.
Voestalpine uses STATISTICA IndustrieProfi- nist as a strategic tool within the framework of industrial data analysis in several areas of the company. The complete equipment of the employees with these common methods is optimized by some Poweruser, which have additional tools such as STATISTICA Data Miner / Process Optimizer and can carry out data analyzes on a more complex level. In doing so, there are regularly important improvement projects, with which the profitability can be considerably increased by looking at the processes as a whole.
No comments:
Post a Comment