Comparison of long-term adaptivity for neural networks

Neural Networks can be used for the prognosis of important quality measures in industrial processes to complement or reduce costly laboratory analysis. Problems occur if the system dynamics change over time (concept drift). We survey different approaches to handle concept drift and to ensure good prognosis quality over long time ranges. Two main approaches - data accumulation and ensemble learning - are explained and implemented. We compare the concepts on artificial datasets and on industrial data from three cement production plants and analyse strengths and weaknesses of different approaches.