Constrained Nonlinear Control Allocation based on Deep Auto-Encoder Neural Networks

Abstract

An intelligent control allocation method for modern reentry vehicles with a strong aerodynamic nonlinearity is presented in this paper. A specially designed deep autoencoder (DAE) neural network is proposed that shares similar information flow with the control allocation process. This similarity is showed by setting the decoder to approximate the control effector function, and by setting the encoder to allocate the expected control moments. The decoder is trained independently based on the aerodynamic coefficients database, and with the help of the well-trained decoder, the encoder is then trained in an unsupervised way without labeled data. This proposed control allocation method could deal with strong nonlinearity of the control effectors at a high accuracy, thanks to the powerful modeling and regression ability of deep neural networks. Numerical examples are provided by the end that explain the training and implementation details, as well as the strong learning and modeling ability of the deep neural network.