Design a Ship Autopilot using Neural Network

Ship autopilots play an important role in insurance of safe navigation and efficient transportation as else. For their successful design and development, many control techniques were and are being developed. In this paper, the application of artificial neural network (ANN) is investigated to design an autopilot for a surface ship. Feed forward multilayered architecture of ANN is used for approximation of the inverse model of the ship. The designed autopilot acts like an optimal one because of using a cost function for generation of ANN training data. The performance of designed autopilot is evaluated in still water and different wave frequencies. The stability and robustness of the designed system is proved through simulation, carried out in Matlab. The results show that the proposed autopilot can efficiently be used to control the course of a ship in a range of parameter variation. 1. Introduction The main duty of an autopilot is to keep the course of ship in a predefined direction using mostly the rudder. The safety and efficiency of transportation are directly dependent on the performance of ship autopilots. For this, a lot of techniques like proportional-integral-derivative controller, optimal control theory, adaptive control, and nonlinear control are used to improve the autopilot performance. The major feature of these approaches is that they require the exact knowledge of dynamics of controlled objects, which is difficult to obtain in practice due to complexity of ship hydrodynamics. For adaptive and nonlinear control there are some demerits like difficult design and stability analysis. The mentioned above disadvantages can be overcome by using artificial neural network (ANN) because of ANN approximation ability of arbitrary smooth function with required accuracy (Haykin 1999; Terekhov 2002). The first way of application of ANN to ship steering is to use a conventional controller for training the neural network (NN) controller (Endo et al. 1989; Burns & Richter 1996; Unar 1999; Khizer et al. 2013; Pathan et al. 2012a, 2012b; Velagic 2006; Zirilli et al. 2000; Alarcin 2007; Alarcin et al. 2010). Another way to develop NN controller is to integrate NN with other control techniques in sense of adaptive control (Cao et al. 2000; Tzung-hang et al. 2001; Ming-chung & Zi-yi 2013; Li et al. 2004; Jun & Liu 2013; Xu et al. 2014; Xiao et al. 2011; Wang et al. 2013; Yang & Zhao 2006; Leonessa & VanZwieten 2004; Hu & Pan 2012).