Ann-based metamodelling with clustering of output values as an approach to robust inverse analysis

by

Ann-based metamodelling with clustering of output values as an approach to robust inverse analysis

Grzegorz Gorecki, Łukasz Rauch, Maciej Pietrzyk, Jan Kusiak

AGH University of Science and Technology, Department of Applied Computer Science and Modelling, Al. Mickiewicza 30, Krakow, Poland.

DOI:

https://doi.org/10.7494/cmms.2014.3.0488

Abstract:

Inverse analysis, which uses artificial neural networks as direct problem models, is gaining popularity. To improve robustness of the metamodel the idea of clustering of the networks based on their output was considered in the paper. This idea splits data by the output values, which in the inverse analysis are known as a part of the objective function. The individual networks are trained at small ranges of output values and they are supported by general networks trained on a wide range of data using special maps. The maps indicate which small network should be used to obtain more precise results. Possibility of using the general wide trained network for checking after the inverse analysis, whether the network for small range was correctly selected, is the main advantage of this method. The basic principles of this approach are described in the paper. Case study for identification of material flow stress model confirmed very good capabilities of this technique.

Cite as:

Gorecki, G., Rauch, Ł., Pietrzyk, M., & Kusiak, J. (2014). Ann-based metamodelling with clustering of output values as an approach to robust inverse analysis. Computer Methods in Materials Science, 14(3), 167 – 179. https://doi.org/10.7494/cmms.2014.3.0488

Article (PDF):

Keywords:

Inverse analysis, Artificial neural networks, Metamodel, Clustering by output values, Material flow stress

References: