WISCON-Methodenentwicklung fuer intelligentes Monitoring Abschlussbericht

At present there is a research deficiency in dealing with processes in environment and economy which cannot or only partly be modelled mathematically. Therefore, the aim of this project was to develop and apply machine learning algorithms for diagnosis and control of such kind of processes and their integration into an intelligent monitoring and control system. Based on the interpretation of measured values (process parameters or time series) process state analysis, cause and effect-analysis and automatic generation of control programs were carried out. Principles of symbolic and subsymbolic processing were combined and thus, both AI-methods as well neural network approaches exploited. Empirical and model knowledge were used if being available. Following learning algorithms have been developed: 1. learning classifications of the current process state, based on a classifier, which generalizes measured and preclassified process states serving as training data set. This leads also to an automatic synthesis of optimal control processes. 2. interpretation of measured time series by their holistic classification with a possible transformation into symbol strings and applying graph metrics. The learning algorithms are both implemented as stand alone versions and integrated into a special version of the monitoring and control system PROVIS especially provided and tested for the WERRA application (prediction and control of high water events). CAL5 and DIPOL can be applied to classification, prediction and process control in other domains, too. Within the ESPRIT-project Statlog they were tested and, in comparison to other learning algorithms, came out very well. The TSC package is a new approach in the field of time series classification and especially suitable for curve classification, automatic extraction of features characterising curves and forecasting tasks. (orig.)SIGLEAvailable from TIB Hannover: F95B2234+a / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekBundesministerium fuer Forschung und Technologie (BMFT), Bonn (Germany)DEGerman