An efficient GPR-based system identification method for UUVs with feature selection via a deep partial auto-encoder

In this paper, an efficient Gaussian process regression (GPR) system identification method is proposed for unmanned underwater vehicles (UUVs) via a deep partial auto-encoder (DPAE). A small set of input features is selected to represent most of the discriminative UUV features as GPR inputs without the loss of prediction accuracy. The DPAE is designed to extract low-dimensional features from a large amount of features of an UUV. Characterized by an output dimensions smaller than the DPAE input dimensions, this design contributes to minimizing the DPAE’s root mean square error (RMSE) in the loss function and reduces training time. Additionally, a novel loss function is proposed to optimize the DPAE’s weights, which incorporates both group Lasso regularization and the GPR prediction error within the loss function. Finally, the particle swarm optimization (PSO) is used for GPR hyperparameters optimization, improving the prediction accuracy compared to the gradient descent (GD). The models of a simulated UUV and a real UUV are identified using the proposed method. Results show that GPR with low-dimensional input features can accurately predict the UUV state and enhance prediction efficiency.