ニューラルネットワークに基づく逆問題解析法のための学習データ適切化法

抄録

This paper proposes a new regularization procedure for inverse analyses using hierarchical neural networks and computational mechanics. The present regularization method, named here as the generalized-space-lattice (GSL) transformation, transforms localized learning data points onto lattice points in a generalized multi dimensional lattice space. This procedure improves the geometrical structure of the learning data sets. The method of inverse analysis using the hierarchical neural networks with the GSL transformation consists of the following three phases : (1) preparation of learning data, which are produced through many finite element computations and then converted by the GSL transformation, (2) training of neural network, and (3) utilization of the trained neural network. Fundamental performances of the GSL transformation are clearly demonstrated through its application to the structural identification of a vibrating non-uniform beam in Young's modulus.