Ship Motion Identification And Prediction Based On Deep Learning

There is no doubt that ocean transportation is of great importance,as more than two-thirds of global freight is transported by sea.The rapid development of the shipping industry is driven by the application of a large number of intelligent technologies,especially in the field of smart ships.The identification of ship systems in the field of intelligent ships can improve system identification accuracy for manipulability controller design.Ship motion attitude prediction can be useful for calibration of intelligent facilities on board such as video acquisition windows.Deep neural network methods have strong non-linear feature capturing capabilities that are suitable for high-precision modeling and prediction work.This thesis proposes a Convolutional Neural Network(CNN)-Long Short-Term Memory(LSTM)network model based on the Attention Mechanism to identify the dynamics of ship maneuverability models.This model is used to identify high-order nonlinear terms in simplified linear maneuverability dynamic models.For the ship motion attitude prediction problem,a multi-channel ship motion state time series data-based prediction model integrating attention mechanism and CNN-LSTM networks is developed to predict ship rolling motion.This model construction is specifically aimed at improving the accuracy of ship rolling motion prediction in terms of wave resistance,and the specific contents were as follows:(1)Identification of ship maneuvering motion mathematical modelThe simplified linear maneuvering model identification is carried out using the system steady-state motion data to construct the system steady-state constraints.Multiple steady-state constraint equations are constructed based on constant input tests,including rudder tests and zigzag tests.In order to determine the system dynamic parameters,coefficient fitting is performed using multiple sets of dynamic input test data,including acceleration-deceleration tests and zigzag tests.The simulation data of a large container ship is processed to meet the requirements for identifying the mathematical model of ship motion.On the basis of the simplified linear maneuvering model,the high approximation and pattern extraction abilities of deep neural networks are used to extract the high-order nonlinear terms of the dynamic model.Based on the simplified linear dynamic model constructed according to the system steady-state constraints,a convolutional neural network is used to extract state and control input data and map them to higher-dimensional feature maps.The attention mechanism is used to capture the importance weights of different channels of the multi-channel input data for the target output value.The LSTM layer is used to extract longterm time series features and model the data.(2)Prediction of ship seakeeping motion attitudeA CNN-LSTM network model with attention mechanism is constructed to predict seakeeping motion.The research object of this thesis includes large oil tanker test data from navigation simulator,KCS ship model towing tank test data,KVCLL2 standard ship model test data,and simulated ship test data.These four types of data sets are divided into training data and test data.Based on the training sample data set,the CNN-Attention-LSTM prediction model is trained,and the trained prediction model is used to predict the ship’s wave-resistant attitude in the test set.The research results show that based on the system steady-state constraints,a simplified linear dynamic model is constructed in this thesis.Moreover,the nonlinear high-order term error identification of the simulated container ship is achieved by combining attention mechanism with the CNN-LSTM model organically.Secondly,a CNN-LSTM prediction model with attention mechanism is proposed,the seakeeping motion attitude of the simulated large tanker,KCS ship model,KVLCC2 ship model,and simulated naval vessel is predicted,and the proposed combined prediction model is compared with the actual error results.The results show that the combined prediction model has better predictive performance.