Ein Verfahren zur automatischen Erzeugung intelligenter Prozessüberwachungssysteme

Manufacturing sites in developed countries can only exist in the long run if their production processes are automated. Prerequisite for this automation is the existence of reliable monitoring systems to detect and fix unfavorable process states. Monitoring systems capture sensor data from the observed process. The current process state is deduced from this data, either based on fixed rules (thresholds, envelopes...) or by the means of artificial intelligence. Intelligent process monitoring systems are considered to be the more powerful type. Currently the creation of intelligent process monitoring systems for a given manufacturing method is time-consuming and requires a high degree of expert knowledge. This is a major barrier for a comprehensive application of such systems. This thesis presents a method for the automatic creation of intelligent process monitoring systems for arbitrary cyclic production processes. These production processes can be described by a set of training data. Each data record in this training data set contains both: data that has been measured in a single cycle of the production process, and the desired prediction result for this production cycle. Based on these training data, the new method generates a process monitoring system that is able to assign predictions even to such data records that have not been part of the training data set. For the realization of the proposed system, a generic process monitoring system was designed and implemented. This system provides the infrastructure for data acquisition and the data streams required for the generated monitoring systems. In the first stage, the generic system does not hold any program logic for processing and evaluating the acquired data. This logic is provided by an external analysis model. Such a model is a processing chain integrating methods for signal preprocessing, feature generation, feature selection and classification. By setting the analysis model, the generic process monitoring system can be adapted to any manufacturing process. With the concept described above, the creation of a process monitoring system for a manufacturing process can be reduced to an optimization problem. The goal is to find the analysis model that adapts best the generic monitoring system to the given manufacturing process. To solve this optimization problem, a heuristic optimization algorithm named Artificial Bee Colony Optimization is applied. For the method proposed in this thesis, the original Artificial Bee Colony Optimization was adapted to handle non-real valued problem spaces