Course Keeping Control For Ships Based On Robust Neural Network

The ship motion has large inertia characteristics, the time const is tens seconds, even more to several hundreds seconds, the rudder responses slowly, the ship motion is nonlinear and time varying. The steering engine that executes motion control exists nonlinear characteristics, for example saturated zone, dead zone, and so on. Ship's parameters variations cause model perturbation; moreover, the running environment is complex. These factors influence ship, and make it deviate from its setting course. Course keeping for ships not only guarantees the ship's safety, but also supports the research on track keeping, dynamic positioning and automatic avoidance collision. In order to solve these problems, it has the theoretical and practical significance that robust neural network is used for course keeping for ships.This thesis integrates robust control with neural network control. Firstly, direct control of neural network is trained to be the inverse of controlled plant; secondly, the generalized controlled object is composed of the inverse model and the controlled plant; finally, robust controller is designed by closed-loop gain shaping algorithm. The controlled plant is similar to a unit matrix. Neural network is adaptive to the nonlinear and time varying plant; robust controller can guarantee robustness of the designed system. BP neural network has five layers structure. In this thesis, convergence rate of BP network is improved by conjugate gradient method; convergence time is quickened, training efficiency is enhanced by combining intensive training with moderate training. The method of signum function is used approximately, while errors propagate through BP network for parameter adjustment. It gains the better control result than before, on the basis of guaranteeing the causal relationship between input and output. Robust controller is designed by closed-loop gain shaping algorithm, this method has the advantages of simple design procedure and obvious physical sense.In VC6.0 environment, simulation experiment is carried out to "Yang Cheng Hu" by C language. The method may keep setting course without overshoot for nominal model. The overshoot for course keeping is controlled within 5% without static error under severe sea states and for±15% model perturbation. Adjusting time is 300s or so.The robust neural network control algorithm can realize course keeping in all kinds of environment by direct control, controller appears satisfactory adaptability and robustness. In addition, the modular design makes the algorithm improvement, the functional addition convenient by C language, is beneficial to applying theory to engineering. This method is the base that prototype of high-performance ship autopilot is made.