关键词: 神经网络, 降维方法, 主成分分析, 自动编码器, 特征提取和融合, 参数反求

Abstract: In dealing with problems of inverse design, optimal design, reliability design and analysis of complex structures, high dimensionality of the datasets can result in high computational difficulty, low computational efficiency and accuracy. Proposes an adaptive dimensionality reduction method for high-dimensional data——Guided Constraint Autoencoder. This method proposes a dual strategy of orthogonal bootstrapping and orthogonal constraint for unsupervised training of autoencoder networks based on the ability of principal component analysis to obtain high quality linear low-dimensional data based on the correlation of high-dimensional data and the unsupervised training of autoencoder with self-reconstruction of dimensionality reduction data quality to achieve high accuracy and high efficiency of dimensionality reduction. The method primarily employs the following techniques: PCA is used to quickly create a projection matrix with orthogonal properties from feature vectors, which is then used to preset the trainable parameters in the autoencoder to guarantee the orthogonality of the initial training parameters; The weight orthogonality constraint (uncorrelated feature constraint) is imposed in the next training to reduce the redundancy of information and ensure the independence of the low-dimensional data. At the same time, a unit-parameter constraint is imposed to avoid gradient disappearance. These strategies result in dimension reduced features with orthogonal properties. To verify the effectiveness of the GAE method, the method was applied to the material parameter inversion of carbon brazing and glass fibre composite laminates, and the performance of different dimensionality reduction methods was evaluated by the accuracy of the parameter inversion. The results show that the GAE method can efficiently obtain high-quality linear and non-linear dimensionality reduction information, effectively overcoming the shortcomings of the principal component analysis dimensionality reduction method which can only be used for linear reduction of high-dimensional information and the low accuracy of standard autoencode dimensionality reduction data, and significantly improving the parameter inverse accuracy of composite laminates.

Key words: neural networks, dimension-reduction, principal component analysis, autoencoder, feature extraction and fusion, parameter inverse