Crystal Ball : die Gewinnung von verwertbarer Information aus Datenobjekten mit unscharfem Zusammenhang

In times of globalization opportunities but also challenges are rising for the enterprises. The permanent competition requires continuous improvement of products and processes from all companies. The companies can set themselves apart from the market by better quality. Beside the increase in customer satisfaction, improvement of quality leads to reduced costs by higher reliability in planning and reduced scrap. A rising competiveness of the company is the result.
The production of parts for the automotive industry has been faced with very high quality demands for a long time. Taking the semiconductor industry as example, the state of the art will be described.
Reproducing single process steps -which have been defined as correct before- as identical as necessary is the main target of tools and procedures which are currently used. Especially in very complex production processes, the impact of variations in a single process step for the final product is often not clear.
Many companies -from small ones to the very big ones- made a lot of efforts together with universities and institutes. They searched for solutions to deal with the huge amounts of data and they tried to gain useful information from this data.
The focal point of this work is the retrieval of root causes for deviations in a very complex production environment. Because of the huge amount of influencing parameters and the long production time, an analytical solution does not lead to the target.
As described in the literature, Neural Networks are often used for classification purposes or for prediction of the behavior of complex systems. In this elaboration, a Neural Network is used to find the root cause for the system behavior. Also new is data reduction by Feature Selection, whereas the Neural Network acts as validation instance for the selected parameters. No knowledge about the technical dependencies is being used in this process. So, no possible solutions are excluded in advance.
With this combination of data reduction (Feature Selection) and knowledge generation (Neural Network), the relevant input parameters for the analyzed effect can be identified successfully. Furthermore, the rules of the Neural Network provide hints for process optimization and error prevention.
At first, this newly developed method has been validated successfully with synthetic data. After that, the system has been used with real production data obtaining convincing results.
Analyzing one problem as an example, more than 1000 datasets from one product were selected. The system identified a small number of parameter responsible for the effect from a bunch of nearly 100.
Actually, the method is being implemented at a semiconductor manufacturer. Other companies show interest in this method.
Beyond that, an implementation of this method seems to be reasonable in other parts of the economy, especially when there are business processes with a combination of high quality requirements and a very good documentation of these processes.