On the optimization of training parameters of BP network for modeling 3D woven composites

Artificial neural network is applied to build a mapping relationship between 3D woven reinforcement and properties of composites. To facilitate the construction of the network, a modified learning strategy for training non-linear network models is developed, based on network system error gradient, with both incremental and decremental factors of the learning rate being adjusted adaptively. The golden section law is put forward to optimize the network training and a series of data from existing 3D woven composite samples is used to train and test the network parameters. By evaluating the network performance in respect to convergent speed and prediction precision, the effectiveness of the proposed learning strategy is then illustrated.