Deep learning machine based ship parametric rolling simulation and recognition algorithms

In this work, we propose a "grey box" algorithm consisting of a deep learning network and parametric roll equation to provide the parametric roll simulations. A basic unsupervised method--principal component analysis is introduced into this algorithm to improve efficiency and accuracy. Besides, it has been proved that different wave time histories generated by the same wave spectrum will induce very different parametric rolling behaviors. Thus, we also proposed an algorithm that combines the deep learning method with spectral analysis to recognize the wave time histories that induce large amplitude rolling. A convolution layer and a pooling layer are added to the recognition network to reduce the input dimension. Besides, according to our study of training methods, we have found that extreme learning machine performs much better than back propagation network. Using the principal component analysis technique and the "grey box" based extreme learning machine, we can obtain very accurate simulations of the parametric roll of the C11 container ship in a short time. Besides, by using the spectral analysis and the extreme learning machine with convolution and pooling layer, we can accurately recognize the time histories induce large-amplitude parametric roll of the C11 container ship.