Research On Extreme Short-term Prediction Method Of Ship Motion Attitude Based On Gaussian Process Regression

Due to the vigorous development of the navigation industry,the types of maritime operations are becoming more abundant.Many tasks urgently require very short-term prediction technology for ship movement,such as carrier-based aircraft taking off and landing,launching and recycling of unmanned boats,installation and use of ship weapon systems,etc.This technology has become a hot issue in the field of shipbuilding and ocean engineering.This paper takes the carrier-based aircraft landing guidance system as the application background,and aims to improve the safety of the carrier-based aircraft landing.The focus is on the very short-term prediction method of the ship’s motion posture.The main work of this paper is as follows:First of all,in view of the existing very short-term prediction methods of ship motion attitude almost only focus on the prediction of attitude values,and rarely consider the problem of the uncertainty of the forecast results.This paper introduces the Gaussian Process Regression(GPR)algorithm to the ship motion.In the issue of very short-term attitude forecast,a GPRbased forecast model is proposed,which can obtain interval forecast results.In order to design the most suitable kernel function for ship motion posture sequence for GPR modeling,this paper analyzes the structural characteristics of different kernel functions and the autocorrelation function of ship motion posture sequence,and proposes a kernel function design scheme for ship motion posture sequence,and based on different types of kernel functions to carry out prediction experiments to verify the measured ship motion posture.Secondly,in view of the low prediction accuracy of a single GPR model,this paper proposes a very short-term prediction algorithm for ship motion posture based on Long ShortTerm Memory(LSTM)network and GPR when the training data set is sufficient.(LSTM-GPR),using the idea of two-step prediction,successfully combining the advantages of LSTM and GPR,and the obtained prediction results can obtain more reliable interval prediction results without losing the prediction accuracy of the LSTM model.Through prediction experiments on the measured ship attitude data under different motion states and different degrees of freedom,the effectiveness and superiority of the proposed algorithm are successfully verified.Finally,in the face of practical applications with less continuous available data,it is difficult to train a neural network-based hybrid forecast model with good prediction performance.This paper proposes an Empirical Mode Decomposition(EMD)based on Empirical Mode Decomposition(EMD).and GPR prediction algorithm(EMD-GPR).From the perspective of data preprocessing,using EMD decomposition to obtain a series of independent Intrinsic Mode Function(IMF)components that can highlight the local characteristics of the data and a residual Then,GPR modeling and prediction are performed on each component,and then the prediction results of each component are summed and the final prediction result is output.Through the analysis of the stationarity before and after the decomposition of the measured ship motion posture sequence,it is verified that the EMD processing can effectively reduce the non-stationarity of the sequence.In order to explore the different number of training samples,which forecast method is more suitable,the training samples The quantity is a variable for the prediction experiment,which verifies the superiority of the EMD-GPR-based very shortterm prediction algorithm for ship motion posture proposed in this chapter when the training samples are limited.