Methoden der künstlichen Intelligenz für das Problemmanagement in der Vorserienproduktion

As a phase of product development process, pre-series production includes the validation and verification of preceding planning efforts as well as the handling of any unplanned events with reference to downstream process phases. The transition from virtual product to physical product is particularly critical to success in this context. The initial physical orchestration of all planning elements from development, production engineering and material steering for the purpose of product and process assurance itself induces problems of varying scope. A survey conducted in this regard shows that, according to their own assessment, industrial companies don't succeed in holistically eliminating possible problems during product development. Increasing demands in terms of product complexity and the dynamics of the markets are making it increasingly difficult to achieve an optimum level of quality from a business point of view. It is even more likely that the share of quality-related costs will increase by 30% if companies do not adapt their quality work to current and future requirements. For the 100 largest industrial companies worldwide, potential losses of USD 215 billion are forecast in this respect. Furthermore, it can be assumed that the increasingly induced product and process complexity results in a higher dimensionality as well as a higher volume of problems, whereby the performance of classical quality management methods reaches a manageable threshold. Therefore, the basic approach of this thesis does not focus on extending the spectrum of manual problem solving methods, but on transferring critical problem solving processes to the machine, as a second instance, with much higher computational capacity. In the approach, the human to method relationship is hence switched into a technology to human relationship. For this purpose, the concept of a case-based multi-agent system was built upon, and two mechanisms of human thinking essential for pre-series production were adapted. On the one hand, the principle of case-based reasoning as a memory mechanism, and on the other hand, that of machine learning as a pattern recognition mechanism. The complementary use of both methods shows a particularly high potential for simultaneously high problem volumes with high knowledge intensity. The required functionality of the developed components of a case-based multi-agent system is derived from recognized models of incident and quality management. A special business focus is on the use of already existing, explicit expert knowledge in existing quality management systems. Using the example of pre-series production in the automotive industry, the utility of developed artifacts is demonstrated both experimentally and in an application context. The effectiveness of machine learning is particularly noteworthy in this context due to the established performance of detection and selection mechanisms.