TY - JOUR A1 - Grote-Ramm, Wolfgang A1 - Lanuschny, David A1 - Lorenzen, Finn A1 - Brito, Marcel Oliveira A1 - Schoenig, Felix T1 - Continual learning for neural regression networks to cope with concept drift in industrial processes using convex optimisation JF - Engineering Applications of Artificial Intelligence Y1 - 2022 U6 - https://doi.org/10.1016/j.engappai.2023.105927 SN - 0952-1976 VL - 120 PB - Elsevier ER - TY - CHAP A1 - Lorenzen, Finn A1 - Schönig, Felix A1 - Grote-Ramm, Wolfgang A1 - Oliveira Brito, Marcel A1 - Neuberger, Heiko A1 - Sobolyev, Alexander T1 - Continual Adaptation of Compressor Maps During Operation on Measurement Data For Increased Performance of Model-Based Applications in Turbomachinery T2 - ASME Turbo Expo 2023: Turbomachinery Technical Conference and Exposition, June 26–30, 2023 Boston, Massachusetts, USA N2 - Higher control engineering concepts, including artificial intelligence and model-based control, are recently becoming more popular in complex industrial applications because they enable a significant increase in efficiency. E.g., a model-based approach can be used to enhance the control and monitoring of several coupled components of a larger turbomachinery train. This requires dynamic models of the components, including all machine characteristics, which may experience a shift in their behavior over the course of their lifetime. Such changes occur due to natural fouling, wear of sub-components or external effects induced by maintenance measures. To overcome this problem, we model machine-characteristic maps with artificial neural networks, which may be used as part of prediction models in a model predictive control unit. If an appropriate data set is available, this allows us to adapt the model to the current behavior of the machine without requiring in-depth knowledge about the underlying physics of this “concept drift”, as it is known in literature. However, the training of neural networks is usually connected with high computational effort while the task needs to be executed in real-time during operation. Furthermore, in real operating conditions, machine sensors can only provide labeled data from the current operating point of the machine, which can be stationary for long operation periods of up to several months. In general, this yields only an unrepresentative data subset of the machine-characteristics, which is not sufficient to retrain the whole model to a new state. To solve these problems, we firstly reduce the model adaptation to a convex optimization problem, which can be efficiently solved in real-time conditions. Secondly, we use a specialized data management system with which we can integrate historical data to supplement the unrepresentative data subset. In this paper we show the application of the outlined method on a compressor map and discuss the advantages and requirements of the method in the context of modeling applications for industrial turbomachinery. KW - compressors KW - Intelligent control KW - Neural networks KW - Modeling Y1 - 2023 SN - 978-0-7918-8697-7 U6 - https://doi.org/10.1115/GT2023-103458 VL - Volume 4 SP - 1 EP - 7 PB - ASME CY - Boston ER - TY - JOUR A1 - Grote-Ramm, Wolfgang A1 - Lanuschny, David A1 - Lorenzen, Finn A1 - Oliveira Brito, Marcel A1 - Schönig, Felix T1 - Continual learning for neural regression networks to cope with concept drift in industrial processes using convex optimisation JF - Engineering Applications of Artificial Intelligence N2 - Process models in industrial applications, e.g. predictive maintenance or automation, are subject to both divergence from the underlying system due to their time-variant nature and to high complexity resulting from a wide operational range being covered. Hence, regression models require high accuracy for the present system state and at the same time need to be valid across the whole system operating space. While accuracy for the current system state can be gained by updating the model on the current data, the overall validity must often be retrieved from historical or design data. We propose a method to find an appropriate compromise for these two demands. A pre-trained artificial neural network (ANN) is continually updated on the current sensor data stream using convex optimisation. Thus, a unique and optimal solution is generated in each update step, while robust regression accuracy on the domain that is not covered by the arriving data subset is maintained. This is achieved by introducing a data management system to provide some historical data, constraining the optimisation problem and manipulating the architecture of the ANN. Models updated with this method show reasonable stability but display plastic behaviour at the current operating point. KW - Continual learning KW - Concept drift KW - Catastrophic forgetting KW - Convex optimization Y1 - 2023 U6 - https://doi.org/10.1016/j.engappai.2023.105927 SN - 1873-6769 VL - 120 PB - Elsevier CY - New York ER -