In this paper, a dynamic Bayesian network (DBN) controller is used to design a ship autopilot. Firstly, the method to build a DBN controller is introduced. Secondly, a simple mass-spring-damping system is dealt with to demonstrate the validity and feasibility of the DBN controller. At last, the DBN controller is applied on a ship autopilot considering wave disturbances. The simulation results demonstrate the validity of the DBN controller showing good performance on the autopilot of ship compared with the traditional PID control method.

Bayesian Network (BN) has gained enormous interest in the research of machine learning, pattern recognition and artificial intelligence (Deventer, et al., 2004). One famous application of BN is medical diagnosis which is called PATHFINDER networks by Heckerman, Horvitz and Nathwani (1992). Kipersztok and Provan (2003) used BN for the fault diagnosis of Boeing commercial aircrafts. BN is also embedded in the user assistant of Microsoft Office products by Microsoft Research. Besides, the fields where BN has been used in include medicine and biology, and so on. In control engineering, DBN has lots of applications as well, in most of these applications, DBN shows its advantages as a good modeling tool or a state observation. Deventer, et al. (2000) considered a new usage of DBN in the control system as a controller, they used DBN to inference the input value of the control system. Hommersom and Lucas (2010) made printing systems adaptive using DBN as decision making. Zhang, et al. (2005) proposed a new method for stabilization of networked control systems with random delays through Markov chains.

BN can be trained by some algorithms from complete data or even incomplete data. The main purpose of BN is to do probabilistic inference. Once one node in the networks gets evidence or desired value, the probabilistic of any other nodes in the networks can be achieved by the Bayesian inference. If one BN can be trained from some historical data of the controlled system, it can calculate the most suitable input value which leads to a desired output value. It can achieve good inference data for the controlled system even the environment or the structure of the controlled object changes, which is the advantage of BN.